TOWARD RELIABLE AND SCALABLE LONG RANGE NETWORKING FOR RURAL IOT

By

Yidong Ren

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

Computer Science—Doctor of Philosophy

2025

ABSTRACT

The Internet of Things (IoT) holds great promise for transforming rural applications such as preci-

sion agriculture, infrastructure monitoring, and environmental sensing. However, enabling reliable

and scalable wireless connectivity in rural areas remains a fundamental challenge due to sparse

infrastructure, energy constraints, and wide-area coverage requirements. This dissertation presents

a system-level exploration into building practical, low-power, and cost-effective LoRa-based net-

works tailored for rural IoT.

Specifically, I address four key challenges: (1) unreliable backhaul due to the absence of cel-

lular or wired infrastructure, (2) weak LoRa signal coverage in complex and obstructed rural ter-

rains, (3) limited scalability of existing backscatter systems for battery-free communication, and

(4) inflexible physical layer encoding that fails to meet the diverse demands of rural applications.

I propose and validate a series of novel techniques, including opportunistic satellite backhaul us-

ing lightweight link estimation and routing, polarization-aligned underground communication for

cross-soil sensing, concurrent non-linear chirp backscatter for scalable battery-free transmission,

and reconfigurable chirp encoding for adaptive coverage, throughput, and energy balancing. These

techniques are evaluated through real-world deployments, hardware prototypes, and empirical ex-

periments across rural-scale testbeds.

Together, these contributions advance the design of robust and scalable LoRa networks for rural

IoT. Looking forward, this work motivates future research on integrating space–air–ground archi-

tectures, embedding joint sensing and communication capabilities, and co-designing cross-layer

protocols that adapt to the dynamic and heterogeneous nature of rural environments. This disserta-

tion lays the foundation for next-generation rural IoT systems that are not only technically efficient

but also practically deployable across underserved regions.

Copyright by
YIDONG REN
2025

ACKNOWLEDGEMENTS

I am profoundly grateful for the guidance, support, and opportunities provided to me during my

journey as a Ph.D. candidate.

First and foremost, I would like to express my deepest gratitude to my advisor, Prof. Zhichao

Cao, for his unwavering mentorship, insightful advice, and continuous encouragement throughout

my research. His vision and dedication have greatly shaped both my academic development and

professional growth.

I am also thankful to my collaborators — Younsuk Dong, Prof. Mi Zhang, and Prof. Shigang

Chen — for their invaluable feedback, inspiring discussions, and collaborative spirit that enriched

my work.

I would like to extend my sincere appreciation to my dissertation committee members, Prof.

Tianxing Li and Prof. Huacheng Zeng, for their thoughtful feedback and guidance during key

stages of my Ph.D. journey.

I feel fortunate to have shared this journey with my lab mates - Huaili Zeng, Yimeng Liu, Gen

Li, Maolin Gan, Jingkai Lin, Guangliang Liu, Bocheng Chen, Shichen Zhang, Fangwei Zhang,

whose friendship and support made EB1100 a vibrant and motivating place—truly the best lab!

Lastly, I owe my deepest gratitude to my family. Their unwavering love, patience, and belief in

me have been the foundation of all my pursuits. This work would not have been possible without

their endless support.

iv

TABLE OF CONTENTS

CHAPTER 1

INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Motivation .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 System Challenges and Technical Limitations . . . . . . . . . . . . . . . . . .
1.3 Contributions and Proposed Techniques . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Organization . .

.

.

.

.

.

.

.

.

1
1
4
6
8

CHAPTER 2

.

.

.

.

Introduction . .

SATERIOT: HIGH-PERFORMANCE GROUND-SPACE NETWORKING
FOR RURAL IOT . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9
2.1
9
2.2 Preliminary and Observation . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 System Design .
.
.
Implementation .
2.4
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
.
2.5 Evaluation .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
.
2.6 Related Work . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
.
.
.
2.7 Conclusion . .

.
.
.
.
.

.

.

CHAPTER 3

.

.

.

.

Introduction . .

DEMETER: RELIABLE CROSS-SOIL LPWAN WITH LOW-COST
SIGNAL POLARIZATION ALIGNMENT . . . . . . . . . . . . . . . . 43
3.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Preliminary and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 System Design .
.
.
Implementation .
3.4
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.5 Evaluation .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
.
3.6 Related Work . .
3.7 Discussion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.8 Conclusion . .

.
.
.
.

.

.

.

.

.

.

CHAPTER 4

.

.

PRISM: HIGH-THROUGHPUT LORA BACKSCATTER WITH
NON-LINEAR CHIRPS . . . . . . . . . . . . . . . . . . . . . . . . . . 76
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.1
.
Introduction . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
.
4.2 LoRa Preliminary . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
.
4.3 System Design .
.
Implementation .
4.4
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
.
4.5 Evaluation .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
.
4.6 Related Work . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
.
.
.
4.7 Conclusion . .

.
.
.
.
.

.

.

CHAPTER 5

AEROECHO: TOWARDS AGRICULTURAL LOW-POWER WIDE-
AREA BACKSCATTER WITH AERIAL EXCITATION SOURCE . . . 98
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

.
5.1
Introduction . .
.
5.2 Preliminary . .
5.3 System Design .
Implementation .
5.4

.
.
.
.

.
.
.
.

v

.
5.5 Evaluation .
5.6 Related Work . .
.
5.7 Conclusion . .

.

.

.
.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
.
.

CHAPTER 6

.

.

.

.

Introduction . .

CHIRPTRANSFORMER: VERSATILE LORA ENCODING FOR
LOW-POWER WIDE-AREA IOT . . . . . . . . . . . . . . . . . . . . . 122
6.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.2 Background and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
. . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.3 Design of ChirpTransformer
6.4 Case I: Network Coverage
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.5 Case II: Network Throughput . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.6 Case III: Network Lifetime . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
.
6.7 Related Work . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
.
6.8 Discussion .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
.
.
6.9 Conclusion . .

.
.
.

.

.

CHAPTER 7

CONCLUSION .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

BIBLIOGRAPHY . .

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

vi

CHAPTER 1

INTRODUCTION

1.1 Motivation

1.1.1 Networking for Rural IoT

Rural and remote areas have become a prominent frontier for Internet of Things (IoT) applica-

tions, including precision agriculture, environmental monitoring, and infrastructure management.

In precision agriculture, for example, farms deploy IoT sensors to monitor soil moisture, weather,

and crop health in real-time, enabling data-driven irrigation and fertilizer management [1,2]. Envi-

ronmental monitoring of wildlife habitats, forests, or water resources similarly relies on distributed

sensor nodes in remote locations to collect critical data (e.g., climate measurements or animal track-

ing) over wide areas. Infrastructure management in rural regions – such as monitoring remote roads,

pipelines, or power lines – also benefits from IoT devices that can report faults or conditions from

far-flung locales. These rural use cases share common communication needs that are distinct from

urban IoT settings. Unlike dense urban deployments, rural IoT nodes may be spread across kilome-

ters of terrain with limited access to mains power or existing network infrastructure. Key network

requirements in these scenarios include:

Long-Range Coverage: IoT sensors in farms or wilderness often reside far apart or far from any

communication tower. Wireless links must reliably span distances on the order of several kilome-

ters, far beyond the range of personal-area or local-area networks [1, 3]. Traditional short-range

wireless technologies – such as Wi-Fi, Bluetooth, and Zigbee – typically cover only tens of meters

and thus cannot economically service devices dispersed over large rural areas.

Energy Eﬀiciency: Devices are usually battery-powered or rely on off-grid energy (e.g. solar), so

the communication protocol must minimize energy usage. Transmit power and signaling overhead

should be low enough to allow multi-year operation on batteries [4, 5]. Frequent battery replace-

ments or recharging is impractical in remote fields or forests, so ultra-low-power communication is

essential.

Affordability and Simplicity: Rural IoT deployments often operate under tight budget constraints

1

and in areas with sparse infrastructure. Solutions must be low-cost in terms of device hardware and

operating expenses. Ideally, sensors should connect without requiring expensive base stations or

licensed spectrum [1, 4, 6]. The network should be simple to deploy and maintain by non-expert

users (e.g., farmers), making proprietary or complex setups less attractive.

These requirements highlight why conventional wireless options are insufficient. Wi-Fi and

Bluetooth, while widely used in consumer IoT, cannot reach the long distances needed on a farm

and typically consume too much power for battery-operated sensors. Cellular networks (LTE, 5G)

provide wide-area coverage in populated regions, but many rural areas lack reliable cellular cover-

age or have coverage only near main roads. Even where available, cellular IoT connectivity entails

high energy consumption and subscription costs for each device, which is prohibitive for large-

scale sensor deployments. Emerging cellular-based IoT standards like NB-IoT and LTE-M improve

power efficiency, but they still depend on operator infrastructure and licensed spectrum availability

that may be lacking in remote locales [7]. This gap in networking options has driven interest in a

new class of long-range, low-power wireless technologies.

LPWAN Technologies and LoRa: Low-Power Wide-Area Networks (LPWANs) have emerged

as a compelling solution tailored to the needs of rural IoT. LPWAN technologies (such as LoRa,

Sigfox, and Narrowband-IoT) are designed to send small data packets over kilometers of range while

keeping devices in a low power state [1,3]. Among these, LoRa (Long Range) has gained particular

prominence as a flexible, cost-effective option for rural applications. LoRa uses a spread-spectrum

modulation in unlicensed sub-GHz bands, enabling link ranges on the order of 10–15 km in rural

open conditions and excellent obstacle penetration compared to higher-frequency radios [1, 7, 8].

Despite a modest data rate (typically 0.3–50 kbps), LoRa networks can support the infrequent,

low-bandwidth telemetry that characterizes environmental sensors and agricultural IoT devices.

Importantly, LoRa devices can operate for years on battery power because of the very low duty

cycle and energy-efficient physical layer [4]. The technology’s use of unlicensed spectrum also

means that farmers or communities can deploy their own LoRa networks without regulatory hurdles

or recurring spectrum fees. These attributes make LoRa a promising foundation for reliable and

2

Figure 1.1 LoRa Networking Architecture.

scalable rural IoT connectivity where traditional wireless networks fall short.

1.1.2 LoRa Networking Architecture

LoRa is often used within the LoRaWAN protocol stack, which defines a complete networking

architecture for connecting IoT nodes to the internet. As Figure 1.1 shows, the typical LoRaWAN

architecture follows a star-of-stars topology [4, 7]: resource-constrained end devices (sensor or ac-

tuator nodes equipped with LoRa radios) transmit data over the air to intermediary devices called

gateways. A gateway is a relatively simple relay that listens to LoRa channels and forward packets

from any nodes in range. Crucially, LoRa gateways are bridges between the wireless LoRa domain

and the internet Protocol (IP) world; they usually have a more powerful processor and an IP back-

haul connection. Multiple gateways can be deployed to provide comprehensive coverage, and end

devices do not need to be associated with a single gateway. The gateways then pass along received

LoRa frames to a central Network Server, often located in the cloud or a central site. This fron-

thaul (node-to-gateway) link uses LoRa radio with Chirp Spread Spectrum (CSS) [9] modulation,

while the backhaul (gateway-to-server) link uses any available IP network (e.g., Ethernet, cellular,

or satellite link) to route data to the server. The network server, together with an application server,

performs higher-layer functions: filtering duplicate packets, managing device sessions and security

keys, and forwarding the IoT data to application databases or user interfaces.

3

LoRa GatewayNetwork ServerSensorsandLoRa nodesBackhaulCSSTable 1.1 Challenges, Limitations and Solutions in LoRa networking for rural IoT.

System Challenges

Technical Limitation

Proposed Solution

Rural area backhaul
(Gateway-Server)

Temporal-Spatial
variant satellite link

Chapter 2: Opportunistic satellite backhaul among multiple
gateways [10]

LoRa wireless coverage
(Node-Gateway)

Complex rural
environments

Chapter 3: Polarization-aligned underground coverage [11]
Chapter 6: Reconfigurable encoder-decoder co-design [5]

IoT battery reliance
(Node)

Limited backscatter
deployment scale

Chapter 4: Concurrent non-linear chirp backscatter [12]
Chapter 5: Drone assisted scalable backscatter [13]

Diverse application
(Node-Gateway-Server)

Inflexible encoding

Chapter 6: Reconfigurable encoder-decoder co-design for
coverage/throughput/energy [5]

In practice, rural LoRaWAN deployments might use a single gateway to cover a local area (for

instance, one gateway with an elevated antenna can cover an entire farm or village). In larger areas,

multiple gateways are deployed to increase coverage and resiliency, all connected back to the same

network server (forming a star topology where gateways are star points that collectively service

many nodes). Gateways are often placed at high points (e.g., on top of buildings, hills, or towers) to

maximize line-of-sight range. Since gateways require backhaul connectivity, they may be installed

where wired internet or at least cellular backhaul is available (such as a farmhouse with DSL, or a

hilltop with cellular signal). In truly isolated areas lacking any infrastructure, alternative deploy-

ment models have been demonstrated: for example, a mobile gateway on a drone or unmanned

aerial vehicle (UAV) can periodically overfly remote sensors to collect their data via LoRa and

then carry the data back to the internet once the drone returns to a connected zone [7]. Similarly,

satellite-connected LoRa gateways (or even satellites acting as LoRa receivers) have been explored

to achieve global coverage. These patterns illustrate the flexibility of LoRa networking – combin-

ing long-range wireless fronthaul with any available backhaul – to bring connectivity to devices in

hard-to-reach places.

1.2 System Challenges and Technical Limitations

Table 1.1 summarizes the major system challenges faced in rural LoRa deployments, the un-

derlying technical limitations.

4

1.2.1 Rural Area Backhaul (Gateway-Server)

While LoRa provides a long-range physical link between devices and gateways, the backhaul

from gateways to cloud infrastructure remains a major bottleneck in rural areas. Most existing LoRa

deployments rely on either wired or cellular IP connectivity, which is often unavailable or unreliable

in remote terrains. Satellite communication offers global coverage but suffers from variable link

quality, high latency, and cost constraints. Moreover, many rural settings experience temporal and

spatial variability in link availability—e.g., satellite visibility may depend on orbital patterns, and

the coverage may fluctuate due to terrain and weather. These dynamics complicate real-time data

forwarding and strain energy-constrained gateway operations. An opportunistic, delay-tolerant, and

energy-aware backhaul strategy is thus essential to maintain consistent service in infrastructure-

sparse environments.

1.2.2 LoRa Wireless Coverage (Node-Gateway)

The theoretical range of LoRa—often advertised as 10–15 km in open environments—can

be drastically reduced in practical rural deployments due to terrain-induced shadowing, vegeta-

tion density, and underground placement of sensors. In applications like precision agriculture or

pipeline monitoring, sensors may be deployed underground or in heavily vegetated areas where

radio signals attenuate quickly or suffer from polarization mismatch. Traditional omnidirectional

antennas and fixed configuration radios are poorly suited to these non-line-of-sight or cross-medium

scenarios. Enhancing fronthaul reliability thus requires both physical-layer adaptations and system-

level awareness of spatial deployment constraints.

1.2.3 Battery Reliance (Node)

LoRa-based IoT devices are predominantly battery-powered, limiting device lifetime and in-

creasing maintenance costs, especially in large-scale rural deployments. Energy-harvesting ap-

proaches, such as solar panels, are weather-dependent and expensive. By leveraging ambient source

carrier modulation, backscatter systems enable battery-less communication at 𝜇W power levels,

extending battery lifetime to tens of years. However, existing LoRa backscatter systems face scal-

ability challenges due to severe signal collisions and costly sources requirement for large-scale

5

Figure 1.2 My contributions in the corresponding LoRa networking stack.

deployment.

1.2.4 Diverse Application (Node-Gateway-Server)

Different rural IoT applications have diverse performance requirements. General monitoring

scenarios prioritize broad coverage across large geographic regions. Industrial sites, such as utility

stations and agricultural processing, require dense sensor deployments. They need concurrent trans-

mission for reliable network throughput. Mobile applications, such as livestock or transportation

tracking, operate under highly dynamic wireless conditions. These scenarios require a fine-grained,

adaptive data rate to enhance energy efficiency. However, standard LoRa uses fixed chirp-based

encoding schemes and lacks flexibility in configuration. This makes it hard to adapt to the needs of

diverse rural IoT scenarios.

1.3 Contributions and Proposed Techniques

To tackle these challenges and limitations, this dissertation presents a set of feasible LoRa network-

ing solutions across the node, gateway, and server stack. These solutions span the APP layer, MAC

layer, link layer, and PHY layer—some targeting specific layers, while others take a cross-layer

design approach. As shown in Figure 1.2, each technique addresses critical aspects of reliability,

scalability, or energy efficiency for rural IoT deployments. Besides system challenges and technical

limitations, Table 1.1 also summarizes the corresponding solutions proposed in this dissertation.

6

App LayerMACLayerLinkLayerPHYLayerLoRaNodeLoRaGatewayServerAeroEcho [12]SateRIoT [9]Demeter [10]ChirpTransformer [5]Prism [11]ChirpTransformer [5]Each solution chapter addresses one or more of these challenges through system-level design, im-

plementation, and real-world evaluation.

1.3.1 Opportunistic satellite backhaul (Chapter 2)

I propose SateRIoT [10] as a novel network architecture with temporal link estimation and spa-

tial link sharing that fully exploits the capability of low-cost low-earth-orbit (LEO) IoT satellites

and ground LPWAN techniques. I introduce a lightweight machine learning model for bursty link

estimation of satellite radio and design a multi-hop protocol with carrier sensing capability to share

the data among multiple ground LPWAN radios. We deploy the system with commercial IoT satel-

lite and LoRa radios in 3×3 km², then evaluate it based on real deployment. It achieves up to 3.3(cid:2)

less energy consumption, 5.6(cid:2) reduction in latency, and 1.9(cid:2) enhancement in throughput.

1.3.2 Underground wireless coverage (Chapter 3)

Using LPWAN to collect data from buried underground sensors offers an efficient and non-

intrusive solution. In such cross-soil links, signal loss arises from both propagation-induced at-

tenuation and polarization misalignment due to varying soil conditions. However, the impact of

polarization misalignment has been largely overlooked. I propose Demeter [11], a low-cost, low-

power programmable antenna circuit design to keep reliable cross-soil communication automati-

cally. My design utilizes a coupler to enable computational polarization adjustment on a COTS

single-RF-chain LoRa radio. We evaluate it on customized circuits and commercial devices in var-

ious environmental conditions. It achieves up to 11.6 dB SNR gain, 4(cid:2) horizontal communication

distance, and 5.5(cid:2) energy saving compared with the standard LoRa.

1.3.3 Non-linear chirp backscatter (Chapter 4)

Existing LoRa backscatter systems suffer from severe signal collisions caused by overlapping

linear chirps. I propose Prism by using different types of non-linear chirps to modulate backscatter

data based on standard LoRa carrier signals. Prism enables multiple orthogonal channels, allow-

ing backscatter tags to transmit concurrently at the same frequency. I designed an energy-efficient

method to accurately shift the frequency of the linear chirps to the non-linear chirps. It is imple-

mented with an ultra-low power tag and standard LoRa carrier and evaluated in real environments.

7

The results show a 6(cid:2) higher transmission concurrency than state-of-the-art.

1.3.4 Drone-assisted scalable backscatter (Chapter 5)

The asymmetric distances between source–tag and tag–receiver pairs require high-cost dense

source deployment. I introduce AeroEcho [13], a drone-assisted asynchronous decoding within the

same non-linear chirp type. I co-design the excitation source and tag with a customized packet

format to enable decoding for multiple tags. And then I propose two aerial routing strategies by

using excitation cells to achieve optimal throughput and symbol error rate for large-scale sensor

deployments. AeroEcho enhances the overall throughput of current backscatter transmission by

5.84(cid:2) and individual tag data rate by 12(cid:2) compared to state-of-the-art approaches.

1.3.5 Reconfigurable encoder-decoder co-design (Chapter 6)

Due to resource constraints, the encoder in existing LoRa systems remains fixed and unmodi-

fied. To expand the design space for data modulation, I introduce ChirpTransformer [5], a reconfig-

urable LoRa encoding framework that harnesses broad chirp features on commercial devices. We

combine software design with hardware interrupts to dynamically modulate data using four custom

features on COTS LoRa nodes without incurring additional energy or latency overhead.. The re-

sults demonstrate a 2.38 (cid:2) increase in network coverage with a neural network decoder, a 3.14 (cid:2)

boost in throughput, and a 3.93 (cid:2) of battery lifetime compared to standard LoRa.

1.4 Organization

The remainder of this dissertation is as follows: in Chapter 2, I discuss the opportunistic satellite

backhaul; in Chapter 3, I discuss the polarization-aligned underground coverage; in Chapter 4, I

discuss Concurrent non-linear chirp backscatter; in Chapter 5, I discuss the drone assisted scalable

backscatter; in Chapter 6, I discuss the reconfigurable encoder-decoder co-design; in Chapter 7, I

conclude this dissertation.

8

CHAPTER 2

SATERIOT: HIGH-PERFORMANCE GROUND-SPACE NETWORKING FOR RURAL
IOT

2.1

Introduction

The Internet of Things (IoT) is revolutionizing the way we understand and interact with the

physical world, enhancing precise agriculture [11, 14, 15], environment monitoring [16–18], and

forest fire prevention [19, 20] among others. Low-power wide-area (LPWA) IoT techniques (e.g.,

LoRa [7, 21, 22], NB-IoT [23], LTE-M [24])1 are desirable to fit the scale of rural areas. Their

networking architecture consists of gateways with backhaul links allowing internet access and many

sensor nodes served by those gateways. However, the lack of urban-area networking infrastructures

(e.g., 5G, LTE, wired networks) in rural areas presents a cost-benefit concern regarding the huge

expenses of establishing new backhaul links.

Recently, LEO satellites [25, 26] demonstrate global Internet access for ground satellite ra-

dios with space links. A satellite radio sends its packets to LEO satellites over space up-links,

and then these packets are forwarded to several territorial ground stations, which are Internet-

connected, through space down-links. Exploring existing LEO satellites to enable direct-to-satellite

(DtS) IoT is a cheaper alternative to establishing new backhaul links on the ground. For example,

SpaceX’s [27] subsidiary Swarm [28]2, a commercial-off-the-shelf (COTS) DtS-IoT provider, pro-

vides global connectivity for $5 per link per month.

SWARM uses ultra-small LEO satellites whose orbit altitude is either 462 kilometers or 510

kilometers. The connections between ground satellite radios and the satellites can be affected by

many factors [29–33], e.g., weather, satellite orientation, relative antenna position between the ra-

dios and the satellites, etc. According to our measurement study (x 2.2.3), the available connections

are sparse during a SWARM satellite passing by. In 80% cases, the total connection time is less

than 20% of the pass duration. Moreover, the data rate of SWARM-M138 [34] radio is as low as

1We implement our prototype with LoRa so use LoRa to represent general LPWA IoT techniques in the rest of the

paper.

2We implement our prototype with SWARM so use SWARM to represent general DtS-IoT techniques in the rest

of the paper.

9

Figure 2.1 An illustration of link-aware SateRIoT with multi-hop data sharing.

1 kbps. The sparse connection time and low link capacity motivate us to rethink the ground-space

IoT architecture that can fully utilize the LEO IoT satellite network to achieve optimal IoT backhaul

throughput while keeping efficient energy consumption and low end-to-end delay. As shown in the

left part of Figure 2.1, existing works [35–38] simply equip a satellite radio on each LoRa gateway

so that the LoRa gateway forwards its collected sensory data to the Internet when a satellite passes

by. However, this intuitive ground-space IoT architecture fails to achieve high performance due to

the following temporal-spatial link challenges.

Challenge 1: Lossy up-links during short-window data pulling. When a gateway is connected

to a satellite, the up-links could be lossy, degrading the energy efficiency of the gateway with a

fixed number of packet transmissions. Specifically, deploying power infrastructure in remote areas

such as farms and forests is challenging, complicating the powering of LoRa gateways [12, 39, 40].

The existing IoT gateway utilizes small solar panels for cost-effectiveness and ease of deployment.

However, due to insufficient sunlight, these small panels become unsustainable in variable weather

conditions such as rain, snow, cloudy days, and nighttime. Therefore, energy efficiency is im-

portant for satellite IoT gateways. The energy consumption of SWARM-M138 [34] radio reaches

7.97 mJ/bit in its Tx mode with 1 kbps, which is 408(cid:2) larger than 0.0195 mJ/bit LoRa SF10 976 bps.

A SWARM satellite transmits data beacons to pull data from ground SWARM radios. When a

10

Existing COTS SystemSateRIoTLossy LinkSatelliteGatewayDisconnectedLinkLinkSharingPacketLink EstimationLink SharingSWARM radio receives this beacon, confirming that a connection is established, it performs 6 sub-

sequent mandatory transmission attempts (i.e., a fixed number in SWARM protocol), irrespective

of how many are received. However, in such a short-term transmission window, we observe that

a SWARM radio can still fail to deliver its packets to the satellite despite having received the data

beacon successfully. With the temporally lossy up-link, the enforced repeated transmissions cause

significant degradation of LoRa gateways’ energy efficiency.

Challenge 2: Diverse connection time among spatial LoRa gateways. Due to the spatial diversity

of LoRa gateways, although they all experience sparse connections, we observe the connection time

is quite diverse among them. In our measurement study, when a satellite passes over four LoRa

gateways, different LoRa gateways can successfully transmit their data to the satellite at different

times. In most cases, only one LoRa gateway is available at a time (x 2.2.5). If a LoRa gateway

only transmits the data received by itself, the LoRa gateway with sufficient connection time and

transmission capabilities will keep idle when all its own data has been forwarded. Considering the

buffered data on other gateways with insufficient connection time, the overall network throughput

is lowered.

To address these challenges, in this paper, we propose SateRIoT, a new network architecture

with temporal link estimation and spatial link sharing that fully exploits the capability of SWARM

backhaul connections, thus enabling high-performance ground-space IoT for rural areas. We esti-

mate the temporal link behavior to determine the optimal number of data packets to save energy

and allow gateways to share their idle connection time with others to maximize network throughput

and lower the end-to-end delay. As shown in the right part of Figure 2.1, besides self-generated

data, each LoRa gateway buffers data shared from all other gateways. When the top-middle LoRa

gateway finds that 𝑘 packets should be transmitted in the current time window, it will immediately

transmit 𝑘 packets from either itself or others. The design of SateRIoT includes three key problems

to guarantee high network performance while reducing extra computation and maintenance costs.

Firstly, the quality of a SWARM link is determined by many factors (e.g., atmosphere, satellite

orbit) [29–33], creating a huge feature space that exposes the challenge of balancing between accu-

11

racy and agility. The accuracy cannot be guaranteed if we only use simple features like weather and

satellite orbit information as inputs. If more link information is counted while a SWARM satellite is

passing, we have limited time for the computation of link modeling and traffic scheduling. In Sate-

RIoT, we choose to guarantee accuracy first while optimizing its agility. Specifically, we design a

bursty link model to depict the temporal link behavior. We estimate the number of successful packet

transmissions in an adaptive transmission window in which the link quality is consistent. To find the

optimal trigger timing that balances between sufficient feature collection and limited computation

time, we propose an Acknowledgement(ACK)-triggered lightweight ensemble learning model.

Second, selecting proper model parameters is not trivial to achieve optimal performance in prac-

tice. On one hand, we need to determine a small group of specific features from many available

ones as the model inputs that derive accurate link estimation. Meanwhile, we need to define an

appropriate time length for each transmission window adaptively. Due to the lossy link nature, a

long transmission window makes it hard to keep link consistency, degrading link estimation accu-

racy. However, due to link sharing among multiple LoRa gateways, short transmission windows

lead to frequent model computation and network data status synchronization, causing increased

energy consumption. Additionally, to ensure a transmission window can be configured in practice,

it must align with the timing of the SWARM data-pulling window. To address the issues, SateRIoT

proposes meticulously utilizing features from the physical layer, COTS satellite IoT protocols, and

environmental information. For the optimal window length, we aim to find the longest window

with consistent link quality to save energy. We adopt an empirical method to identify the optimal

configuration.

Third, in link sharing, a LoRa gateway needs to collect data from others. However, when and

how to collect the data is challenging. Since the connection time is sparse, after a LoRa gateway

sets up a transmission window, we have no time to use energy-efficient tree-based ad-hoc routing

protocols for packet pulling from others [41, 42]. Since the energy consumption of a LoRa radio is

far less than that of the SWARM satellite radio, we enable link sharing with a multi-hop flooding

protocol, which utilizes broadcast to achieve network-wide data consistency quickly. Specifically,

12

when a gateway collects new data from its sensor nodes, it initiates a flooding process, where other

gateways forward the packet until all receive a copy.

In this way, each LoRa gateway collects

up-to-date data from others in a timely manner. In addition, carrier sense backoff is adopted to

prevent packet collisions during concurrent LoRa flooding, improving network consistency and

energy efficiency. On the other hand, it is possible for gateways to establish their transmission win-

dows simultaneously. Although collisions can be mitigated by the LR-FHSS modulation [43, 44]

in SWARM, duplicate packets may still occur in link sharing. To prevent duplicates, we prioritize

self-generated data and shuffle relayed data. We also design beacons to lock transmitting packets

and synchronize queues with the flooding service.

We implement SateRIoT on four gateways embedded with COTS SWARM M138 [34] and LoRa

SX1262 [9] radio chips. We deploy these four gateways in several university farms that cover 9 𝑘𝑚2

areas to evaluate the performance of SateRIoT. Moreover, we have conducted extensive trace-driven

emulation experiments for a large network scale with 12 gateways. Our results indicate SateRIoT

delivers comparable throughput using up to 3.3(cid:2) less energy for each gateway. Additionally, Sate-

RIoT reduces packet delay by up to 5.6(cid:2) and boosts throughput by 1.9(cid:2). Our contribution can be

summarized as follows:

• We empirically measure the characteristics of LEO IoT satellite links with real-world de-

ployment and observe three insights of temporal-spatial link behavior, which demonstrate

the barriers of directly adopting the DtS-IoT to achieve high-performance rural IoT.

• We design SateRIoT, a new networking architecture consisting of well-designed satellite link

estimation and sharing modules to enable network-wide data sharing and link-aware data

transmission, enhancing network performance.

• We fully implement the proposed SateRIoT design with COTS LoRa gateways and SWARM

radios. We extensively evaluate its efficiency with real-world deployment and trace-driven

emulation. The results demonstrate that SateRIoT achieves equivalent throughput with 3.3(cid:2)

13

less energy for individual satellite devices. Additionally, our scheduling scheme reduces

packet delay by up to 5.6(cid:2) and enhances throughput by 1.9(cid:2).

2.2 Preliminary and Observation

2.2.1 LEO Satellite for IoT

LEO satellites operate between 500 km and 2,000 km above Earth, offering advantages such

as shorter latency and more frequent revisits compared to satellites in higher orbits. LEO satellite

communication can provide high-speed broadband internet access globally, like Starlink [45] and

Oneweb [46]. Additionally, LEO satellites can also facilitate low-cost IoT in rural areas with limited

network infrastructure.

Satellite IoT has become an emerging field with selective commercial corporations striving to

provide global IoT connectivity without resorting to expensive high-data-rate satellite communi-

cation systems. These entities have launched satellites to form extensive constellations, enabling

the use of ground-based small satellite radios equipped with advanced long-range communication

capabilities. For example, Sateliot [47] proposes to integrate NB-IoT for satellite communications

into both current and forthcoming 5G infrastructures. Lacuna Space [48] delivers an uplink service

utilizing a LoRa-inspired physical layer. SpaceX’s IoT solution [27] SWARM Technologies [28]

collaborates with Semtech [49] to employ long-range frequency hopping spread spectrum (LR-

FHSS) modulation within the VHF bands. This operates with an uplink range of 148-150 MHz and

a downlink range of 137-138MHz, aiming for exceptionally cost-effective satellite communication

at a 60 USD data fee per year. In this study, we use SWARM satellite radio to evaluate the efficacy

of COTS satellite IoT systems and subsequently craft our design based on these insights. SWARM

with 168 satellites differs quite from traditional high-data-rate LEO satellite networks (e.g., Star-

link with more than 10000 satellites) in terms of protocols, satellite size, and constellation scale.

SWARM links exhibit unique characteristics such as a more sparse satellite distribution, shorter

connection times, lower frequency usage, and reduced network costs [28, 32].

14

2.2.2 Measurement Experiments

To evaluate the network performance of the COTS satellite IoT system, we carried out outdoor

measurement experiments on a rural farm. We employed 4 SWARM Eva Kits [50], each fitted

with an M138 modem [34], and powered them using 18-24V DC solar panels. Over a span of 120

hours, we gathered data from 12 unique locations within a large farm, collecting approximately

3,000 packets and spanning an area of 9km2 in rural farmland. We use infinite traffic to ensure we

can capture the link status of every satellite passing by.

2.2.3 Up-link Connection Time

The propagation environment for SWARM satellites operating in the VHF bands introduces

unique challenges associated with the temporal and spatial link variability [32]. For instance, at-

mospheric elements such as precipitation rates and cloud cover can cause time-dependent attenu-

ation. This attenuation can vary over brief periods, making a substantial impact on link quality.

Meteorological changes, especially seasonal variations, exacerbate these dynamics [31]. Due to

unpredictable conditions, a fixed satellite link pattern cannot be relied upon. On the other hand, the

link’s spatial variability is attributed to differences in terrain and topography. These geographical

features can result in shadowing and multipath effects, which differ based on the elevation angles

and the terrain’s contours [31, 33]. Additionally, satellite orientation is critical, especially in low

Earth orbit, due to its high angle sensitivity. As a satellite progresses along its orbit, its relative

positioning with respect to the ground satellite radio also changes. The elevation angle, defined

as the angle between the ground’s horizontal plane and the satellite’s line-of-sight, influences the

atmospheric path length and, subsequently, the link attenuation [25, 51].

Observation 1: Disconnected links During a SWARM satellite passes by, we divide the passing

duration into several time windows with 1 minute. Given a time window, if a packet is successfully

transmitted, we mark this time window as a connected time window. We define the connection

time ratio as equal to the ratio between total connected time windows and total time windows in

the passing duration to quantify the link disconnected property. According to the collected data

traces, Figures 2.2 shows the results based on different maximum elevation angles and satellite

15

Figure 2.2 Connection time ratio.

passing duration. The results show that the connection time ratio remains consistently low. This

trend holds irrespective of the maximum elevation or the duration of satellite passage. The average

and median values are 14.1% and 11.9%, respectively. The highest observed ratio doesn’t exceed

40%, and in 80% of the satellite passes, this ratio is below 20%. Such findings indicate the satellite

links are disconnected, increasing the transmission latency.

2.2.4 SWARM Communication Protocol

Figure 2.3 depicts the communication protocol within the Swarm satellite IoT system. There are

3 basic packet types: satellite non-data/data beacon, data packet, and ACK (acknowledgement).

Dotted edges represent the reception side and solid edges represent the transmission side. As a

satellite coverage period begins, all the satellite radios awaken from their sleep mode. These satellite

radios then start monitoring space, anticipating the satellite beacon. The received satellite beacon

is the unsolicited message from the overhead satellite that can be classified into two types: non-

data and data beacon. The data beacon informs satellite radios of the satellite’s presence and its

readiness to accept messages from the ground. Receiving the satellite data beacon is a prerequisite

for initiating data packet transmission to the satellite. Non-data beacons are used for downlink traffic

and will not trigger the satellite radio to transmit data. Multiple non-data beacons appear before

data beacons. Both beacon types include information on the Received Signal Strength Indicator

(RSSI), Signal-to-noise ratio (SNR), frequency, timestamp, and satellite ID.

According to the collected data traces, during times without transmissions, the non-data beacons

16

020406080Max Elevation(°)020406080100Connection Time Ratio (%)08020 40 60020406080100Connection Time Ratio (%)Satellite Pass Time (minutes)Figure 2.3 COTS SWARM IoT Communication Protocol.

were sent every 60 s. During each data transmission period, the satellite will increase the frequency

of non-data beacons at the beginning and send a data beacon afterward. The time offset between

two data beacons uniformly varies from 20 s to 52 s. All the data is transmitted between two data

beacons. The median and average time gap between two adjacent data beacons during gateway

transmission is 33 s and 35.2 s respectively.

When data packets are successfully received by the satellite, an ACK is promptly sent back to the

satellite radio, as illustrated by the green block in Figure 2.3. The ACK signifies a successful data

packet transmission. It includes RSSI, SNR, the sent packet ID, and frequency information. All

packets delivered to the cloud correspond with acknowledgements received at satellite radio. If an

ACK is not detected within a receiving window, it indicates a failed packet delivery. There are two

main reasons why a packet may fail to reach the satellite: The first is when satellite radio chooses

not to transmit data because of a high background noise level with RSSI exceeding -88 dBm. The

second is when satellite radio transmits packets, but the satellite cannot detect them, either due to

poor signal quality or interference.

Observation 2: Lossy Links. Satellite radio has a transmission buffer with a queue system called

satellite radio queue, and messages are generally dispatched based on their entry sequence into this

queue. Any unsuccessful packets are then re-queued for another attempt in the next transmission

window accompanied by the following data beacon. In practice, the energy efficiency of satellite ra-

17

Data BeaconACKData Packet……SatelliteTerminal6 transmission attemptsFailedRetransmissionNon-data BeaconTime(a)

(b)

Figure 2.4 Lossy link ratio.

dio transmission performance is often degraded by failed transmissions. This inefficiency is further

exacerbated by retransmissions, particularly when the communication channel deteriorates. “Lossy

link” indicates the link condition when a downlink data beacon can be received while the uplink data

packets may suffer loss in the following transmission windows. Using a software-defined radio, we

detected 6 signal pulses on the uplink spectrum. COTS SWARM fixes the uplink transmission of

6 consecutive packets as the default setting to match its temporal dynamics. COTS radios set this

fixed value because more attempts may result in high chance of transmission failures while fewer

attempts might waste good transmission opportunities.

During the same satellite pass, the lossy link ratio is defined as the proportion of 1-minute

transmission windows experiencing lossy links to the total number of transmission windows during

the pass duration. According to our collected data trace, Figure 2.4 illustrates the lossy links ratio

under different maximum elevation and pass duration settings. We can observe that 71% satellite

radios experience 100% lossy links ratio, leading to significant energy wastage, especially in satel-

lite communication radios with high energy consumption. This motivates us to predict transmission

capability to reduce energy waste.

2.2.5 Geo-spatial Distributed Radios

As mentioned in Section 2.2.3, too many factors can influence the satellite communication link.

The link of different geospatial locations can be variable and dynamic. To better understand the

18

020406080Max Elevation(°)020406080100     Lossy Link Ratio(%)08020 40 60020406080100      Lossy Link Ratio(%)Satellite Pass Duration (minutes)(a) Data transmission performance of 4 gateways in each minute.

(b) Data transmission performance of 2 gateways in each minute.

Figure 2.5 Comparisons of transmitted packet number at each minute between two gateways at
different distances.

link dynamics across multiple locations, we compared the performance of multiple satellite radios

operating concurrently to measure the maximum transmission capability as illustrated in Figure 2.6

and Figure 2.5.

Observation 3: Dynamic spatial and temporal links. We analyzed the performance during each

minute to further understand the packet transmission dynamics, as illustrated in Figure 2.5. The bars

represent the number of packets successfully sent to the satellite during a 1-minute time window. In

Figure 2.5a, the satellite radios of 4 gateways transmit packets alternately over time, with primarily

non-overlapping periods. In most cases, only one gateway can connect to the satellite for one minute.

Figure 2.5b further reveals that the two gateways’ sending periods do not overlap at all, occurring

at different times separately. This suggests that transmission capabilities vary by location and time.

Moreover, we analyzed the overall performance of 4 gateways arranged in a square formation,

19

01020304050Time (minute)0102030PacketsGateway AGateway BGateway CGateway D01020304050Time (minute)0102030PacketsGateway AGateway B(a) 4 gateways with 500 m distance

(b) 4 gateways with 800 m distance

(c) 2 gateways with 2300 m distance

(d) 2 gateways with 2300 m distance

Figure 2.6 Comparisons of throughput with the same satellite pass among multiple gateways at
different locations.

where each gateway was equidistant from its neighbors by either 500 m or 800 m with the same

satellite pass duration. The slopes change over time for 4 gateways, demonstrating the temporal

diversity. In Figure 2.6a, A, C, D exhibit varying leadership in overall transmission rates. Notably,

B and D experience minimal data transmission for about 15 minutes, and C surges ahead after a 62-

minute delay, maintaining its lead thereafter. As the observation period progresses, the performance

gap among A, D, and B significantly increases, with A and B achieving 1.9(cid:2) and 1.5(cid:2) the packets

of D. Conversely, Figure 2.6b shows D’s progressive increase, with B and C alternatively leading to

a similar trend, ultimately achieving comparable throughput. A maintains low data rates, managing

only 54 packets over 100 minutes, 2.8(cid:2) less than D, which significantly falls short of transmission

requirements.

20

020406080100Passing Time(minutes)050100150Packets NumberGateway AGateway BGateway CGateway D020406080100Passing Time(minutes)050100150Packets NumberGateway AGateway BGateway CGateway D020406080100120Passing Time (minutes)306090120150180Packets NumberGateway AGateway B020406080100Passing Time (minutes)306090120150180Packets NumberGateway AGateway BFigure 2.7 The network architecture in SateRIoT.

We also compare the performance of two terminals with 2300 m distance. Figure 2.6c illustrates

that A begins its data transmission 20 minutes after B. And B progressively extends its lead over

time. As a result, by the end of a 120-minute period, B has achieved approximately 2(cid:2) the through-

put of A. Conversely, Figure 2.6d presents a scenario where the performance disparity between the

gateways narrows over the same distance. Although B maintains its lead initially, A catches up by

the 43-minute, matching B’s cumulative data transmission. Subsequently, A overtakes B and sus-

tains this advantage up to the 100-minute, achieving 1.3(cid:2) throughput compared to B. Collectively,

these illustrations underscore the dynamic performances of satellite radio under varying temporal

and spatial scenarios.

Temporal and spatial link dynamic creates varying transmission capabilities among satellite

radios. Some satellite radio may struggle to send packets, while others have surplus capacity af-

ter completing the delivery of packets within their own covered ground area. Thus, the network

throughput is degraded.

2.3 System Design

2.3.1 System In A Nutshell

In SateRIoT, a LoRa gateway is equipped with a SWARM radio to access the Internet through

SWARM satellites. On the other hand, it can use its LoRa radio to communicate with other LoRa

21

Satellite RadioPHY LayerLinkEstimationLoRa RadioData Beacon ACKSatellite InfoLink SharingEnvironmentProtocolPhysical LayerFlooding(Data sharing)Data Queue(Avoid duplicate)Relay Data Beacon Packet NumberWindow LengthFeaturesInputBursty Link ModelLight-weight boostingFloodingBeaconSensorDataUplink Datagateways and collect sensory data from sensor nodes. All LoRa gateways operate in a distributed

manner. Figure 2.7 illustrates the network architecture design of SateRIoT, which consists of three

layers: link estimation, link sharing, and physical layer communication radio.

In the link estimation layer, the core design is the bursty link model (x 2.3.2). First, we build

a lightweight model structure (x 2.3.2.1). The temporal burstiness of a link determines the short

stable duration of the temporal links for data packet transmission. Based on our understanding

of SWARM protocols, we design an ACK-triggered scheme to estimate the link, balancing model

accuracy and agility to make it practical for real-time link estimation. Second, features (x 2.3.2.2)

are selected to be the most informative for accurate temporal link estimation. These features include

information from the physical layer, environmental characteristics, and COTS satellite protocols.

Third, we empirically select the longest window length (x 2.3.2.3) to maintain consistent link quality

while keeping the computation and system overhead low.

To combine the capabilities of multi-gateways, we design link sharing module to improve the

overall network performance by exploring link spatial diversity.

the multi-hop flooding protocol

(x 2.3.3.1) uses the LoRa radio to enable efficient network-wide data sharing of sensory data or

flooding beacon. Priority data queue (x 2.3.3.2) avoid duplicate transmission by setting priority

order among data packets. The data packets in the priority data queue are managed by both the link

model and the flooding protocol.

2.3.2 Bursty Link Modeling and Estimation

According to the observation of temporal lossy link, it suggests the up-link is short-term bursty,

which means the link behavior (e.g., success or failure) is only stable in a varied short-time window.

We design a bursty link model that estimates how many packets can be successfully transmitted in a

stable transmission window. To guarantee the current transmission window is bursty for successful

transmission and collect sufficient information to estimate the number of successful transmissions,

instead of using satellite beacons to trigger a link estimation, we trigger the link estimation after

an ACK is received, indicating a packet has been received at the satellite side. Moreover, we can

collect additional critical features from the packet ACK.

22

Figure 2.8 Boosting based link estimation model.

Specifically, based on our observation in x2.2.4, a gateway attempts data packet transmission

once it has received a data beacon from a passing-by satellite. To estimate the current link quality,

the gateway only adds a probe packet to its satellite radio queue. Once the gateway receives the

acknowledgement of the probe packet from a satellite. We initiate a link estimation process to

determine how many data packets should be set as pending status in a transmission window, whose

time length is predetermined as 𝑇𝑡𝑥, including multiple data beacons. From our measurement study,

both the PDR and SNR remain stable across these consecutive data beacons. We leverage the short-

term stability to perform link estimation for a transmission window to lower the link estimation

overhead. When all pending data packets are transmitted once or the transmission window expires,

regardless of whether all ACKs are received, we dequeue the left data packets from the satellite radio

queue and insert a probe packet into the satellite radio queue again to trigger the link estimation of

the next transmission window.

2.3.2.1 Light-weight model structure

To predict the versatile link within a given transmission window, we use boosting model ar-

chitecture to build an effective, lightweight machine-learning algorithm for link estimation and

scheduled packet prediction. Figure 2.8 illustrates our link estimation model architecture and trans-

mission capability prediction. We utilize the four kinds of features as the input X and try to predict

23

PHYInput FeaturesXMax depth: DTreenumber: t (0 ≤ t ≤ N)Tree N(…)Tree tTree 1𝑭(𝑿)=෍𝒇𝒕𝒌(𝑿)𝒇𝟐(𝑿)……𝒇𝟏(𝑿)Packet 1~k (k= 6×w)𝒇𝑵(𝑿)(…)residualresidualDatabeaconnumber: w…𝑻𝑫𝑺𝑵𝑹𝑹𝑻𝑺𝑵𝑹𝑹𝑻𝑹𝑺𝑺𝑰𝑻𝑫𝑹𝑺𝑺𝑰Weather𝑹𝑻𝒄𝒐𝒖𝒏𝒕𝑹𝑻𝒅𝒆𝒍𝒂𝒚ESP𝑻𝑫𝑬𝑺𝑷𝑹𝑻𝑬𝑺𝑷ENVWindowoffsetPass durationNoiseProtocol1 to k implement classifications where 𝑘 = 6 (cid:2) 𝑤, w represents data beacon amount given the

duration of the current transmission window 𝑇𝑡𝑥. The model 𝐹0„𝑥” is initialized to the logarithm
of class priors, i.e., for all 𝑥, 𝐹0„𝑥” = »log 𝑝1, log 𝑝2, . . . , log 𝑝𝑁 …𝑇 , where 𝑝𝑘 is the proportion of
class 𝑘 in the training set. The tree index in our boosting model is t from 1 to N:

1. For each class 𝑘, compute the pseudo-residuals:

[

𝑟𝑡
𝑖𝑘 = (cid:0)

𝜕𝐿„𝑦𝑖, 𝐹 „𝑥𝑖””
𝜕𝐹𝑘 „𝑥𝑖”

]

𝐹 „𝑥”=𝐹𝑡 (cid:0)1 „𝑥”

𝐿 is the multi-class logarithmic loss. For a dataset of 𝑁 samples, the total multi-class loga-

rithmic loss is:

𝐿 = (cid:0) 1
𝑁

𝑁∑

𝐾∑

𝑦𝑖𝑘 log„ 𝑝𝑖𝑘 ”

𝑘=1
𝑦𝑖𝑘 indicates whether sample 𝑖 is in class 𝑘 (1 if true, 0 otherwise). 𝑝𝑖𝑘 is the model’s predicted

𝑖=1

probability that sample 𝑖 belongs to class 𝑘.

2. For each class 𝑘, fit a new tree 𝑓𝑡𝑘 „𝑥” with 𝑟𝑡

𝑖𝑘 .

𝑓𝑡𝑘 „𝑥” = arg min

𝑓

𝑀∑

(

𝐿

𝑖=1

)

𝑟𝑡
𝑖𝑘 , 𝑓 „𝑥𝑖”

3. Update the model: For each class 𝑘, find the coefficient 𝛾𝑡𝑘 that minimizes the overall loss,

and update

𝛾𝑡 𝑘 = arg min

𝛾

𝑁∑

𝑖=1

𝐿„𝑦𝑖, 𝐹𝑡(cid:0)1„𝑥𝑖” ‚ 𝛾 𝑓𝑡 „𝑥𝑖””

𝐹𝑁 „𝑥” = 𝐹0„𝑥” ‚

𝑁∑

6(cid:2)𝑤∑

𝑡=1

𝑘=1

𝛾𝑡𝑘 𝑓𝑡𝑘 „𝑥”

During the training process, we adjust data or feature sample ratios to randomly selectively use

data and features for each tree rather than employing the entire set. This random feature and data

diversity makes the final model more robust, enabling it to better adapt to various data distributions.

We can effectively reduce the risk of overfitting the training data, thereby enhancing the model’s

generalization ability. Our model balances uneven real data sets by giving more weight to under-

represented categories during training. We also use random sampling to oversample/desample the

data with different labels and balance the class distribution.

24

2.3.2.2 Link features

The input of the link estimation model.

Physical layer features: The successful transmission of the probe packet indicates a satellite data

beacon and an acknowledgement have been received by the gateway. RTrssi and RTsnr indicate the

RSSI and SNR of the received satellite data beacon. TDrssi and TDsnr indicate RSSI and SNR

of the acknowledgement of the probe packet. The RSSI and SNR values represent the physical

propagation property of the current down-link from the satellite to the gateway.

Expected Signal Power (ESP) feature: RSSI and SNR can be distorted by environmental noises.

To focus on satellite signals, we combine RSSI and SNR to generate an ESP [8, 52] value, which

indicates the signal attenuation along the propagation path. The ESP feature is calculated as follows:

ESP = RSSI ‚ SNR (cid:0) 10 log10

„1 ‚ 100.1(cid:1)SNR”

(2.1)

Then we can get RTesp and TDesp with physical layer features.

Environmental features: We include several environmental features as follows: Noise: RSSI of

background noise. This is measured by SWARM radio when no satellite appears. Elevation: The

maximum elevation during a satellite pass can be found on the SWARM website [28] given the

location of the gateway. Passing duration: The passing duration of a satellite is announced on

SWARM website [28] according to the location of the gateway. Transmission Window Offset: The

time offset between the current estimated transmission window and the time that the satellite starts

to pass the area. The relative antenna position between the gateway and the satellite is different

at different time offsets. Weather: Based on the weather released by local weather station, we use

four-levels quantization to define the weather values from sunny to drizzle.

Protocol features: RTdelay: The time delay between the received data beacon and acknowledgement

of the probe packet. RTcount: The number of satellite non-data beacons during a 30-seconds period

before the acknowledgement of the probe packet.

25

2.3.2.3 Link window length

The predicted window length of our link estimation model is significant. Predicting long win-

dows can lead to low link estimation accuracy since only 2% of the minutes have data transmission

modes lasting more than 2 minutes. Consequently, selecting overly long observation windows can

significantly degrade link estimation accuracy. Conversely, shorter windows necessitate more fre-

quent transmission of LoRa beacons, which increases both energy consumption and computational

overhead. Therefore, optimizing the window length is critical to achieving a balance between es-

timation accuracy and system efficiency. To maintain consistency and continuity with the trans-

mission patterns, the window length must align with the natural data transmission cycles. Higher

link quality and more reliable data transmission are indicated by consecutive data beacons. Inap-

propriate window lengths can disrupt these continuous transmission periods, resulting in a marked

decrease in estimation accuracy. Thus, careful selection and tuning of window lengths are essential

for improving link quality assessments and ensuring efficient network operation.

2.3.3 Link Sharing

2.3.3.1 Multi-hop Packet Flooding Protocol

The multi-hop flooding protocol provides a primitive networking method to enable network-

wide packet sharing. The basic idea is that a LoRa gateway initiates a flooding process once it

has a packet to share. There are two types of packets: data packets and flooding beacons, with the

latter used to manage the data queue. A LoRa gateway immediately forwards the packet once it

receives one from another LoRa gateway. A flooding packet starts from the original gateway. In

the first round, the packet will be delivered to the closest gateways. Next, these gateways will keep

relaying this packet to their next-hop adjacent gateways. Eventually, the packet can be delivered

to all gateways [53, 54]. For IoT data collection systems in rural areas, flooding protocol is easy

to deploy and implement in the low-cost IoT gateways without coordination overhead, avoiding

complex network traffic patterns.

Carrier-sense based Collision Avoidance.

It is possible that several LoRa gateways initialize

multiple flooding processes in a short period. Additionally, several LoRa gateways may receive the

26

same packet and start to forward simultaneously. Without noticing others’ packet transmission, the

potential packet collision could degrade the reliability of the data sharing. It will be worse in dense

network deployment with larger flooding scale. To solve this, We adopt channel activity detector

(CAD) on COTS LoRa radio chips [55] to enable low-cost carrier-sense-based collision avoidance.

Before a LoRa gateway forwards a received packet, it will wait for a random initial backoff, then

repeat carrier sense until the channel is clean.

Flooding Beacons. The link model of a gateway triggers two types of beacons to synchronize the

data packets among all gateways with the flooding protocol. Firstly, when the gateway determines

which packets will be transmitted in a transmission window, it sends out a beacon including in-

formation on these packets. Secondly, when the transmission window ends, the gateway sends a

beacon indicating which packets have been successfully transmitted.

Network Consistency. In case a LoRa gateway misses a packet from other gateways due to LoRa

link dynamics [8]. Each LoRa gateway maintains the status of its local buffered packets and broad-

casts the status with the schedule of a Trickle timer [56]. If a gateway receives others’ status and

finds an inconsistency with its local status, it will request other gateways to send the missing packet.

The trickle timer will be reset when a gateway receives a request. This strategy offers both a system-

atic approach to maintaining information consistency and a proactive method to resolve potential

mismatches in a distributed environment. In this way, all LoRa gateways consistently buffer all

packets and receive the flooding beacons from others for later global packet transmission schedul-

ing.

2.3.3.2 Priority Data Queue Management

Priority and data enqueue: The priority data queue structure includes a self-generated data

queue and a relay data queue. The self-generated data queue has higher priority than the relay data

queue. Since self-generated data packets are unique across different gateways, duplicate packets

will be prohibited when multiple gateways simultaneously send data packets to a satellite with

non-collided frequency hopping. We do order shuffling for the relay data queue. Namely, when a

gateways received a relay data packet, it will insert the data packet to a random position in the relay

27

Figure 2.9 Outdoor deployment of satellite IoT radio with different weather.

data queue. In this way, when multiple gateways concurrently transmit data packet, the number of

duplicate packets will be further reduced since they will forward the relayed data in different order.

Packet holding and releasing When a gateway receives a flooding beacon from others indicating

the data packets they will transmit in the coming transmission window, the gateway will check its

queue. For those identical data packets, the gateway will hold them for a holding period that equals

two transmission windows 2𝑇𝑡𝑥. The holding packets have no chance scheduled if a transmission

window is coming. In the holding period, if the gateway receives another flooding beacon (x2.3.3.1)

indicating the holding packets are successfully transmitted, it will dequeue them directly. When the

holding period expires, the gateway will release them. In this way, we actively reduce potential

duplicate transmissions. For each gateway, when a transmission window ends, it will dequeue

packets whose ACKs have been received.

2.4

Implementation

Our methods are fully compatible with LoRaWAN and COTS satellite devices without addi-

tional hardware or centralized coordination. The data for the SWARMIoT only costs 5 USD each

month [28], much less than the cost required for network infrastructure construction in rural areas.

Ourdoor SWARM satellite radio deployment: We employed 4 Swarm Eva Kits [50], each fitted

with an M138 Modem [34], and powered them using 18-24V DC solar panels. Figure 2.9 shows

the outdoor Swarm ground gateway deployment scenarios to collect data in four different weather

28

SunnyMostly CloudyPartly CloudyFull Cloudy orDrizzleSWARM M138 Radio18-24V Solar Pannel VHF AntennaFigure 2.10 Outdoor deployment of flooding experiments and satellite radio locations.

conditions (e.g., sunny, partly cloudy, mostly cloudy, fully cloudy or drizzle, rain). Two simi-

lar weather conditions can coexist during the same time period in a close area. The locations of

SWARM devices are shown as orange circles in Figure 2.10 around 9 𝑘𝑚2 rural area farmland. For

the satellite radio, the length of a packet is 192 bytes that can contain multiple sensory data packets.

Outdoor LoRa radio deployment: Figure 2.10 depicts the outdoor deployment of LoRa radios in

9 𝑘𝑚2 farm zones. The yellow triangles represent locations. We deploy LoRa devices in elevated

positions with open space, such as high fences, small trees, or on tripods at heights ranging from

1.5 m to 2.5 m. We use 4.2 dBi gain LPWA antennas [57] and 12 SX1262 [9] radios controlled by

ESP32 MCUs [58] operating on US915 ISM bands with SF12 and 125kHz bandwidth. We uses

LMAC-1 [55] to enable CSMA for LoRa transmission.

Link estimation model: The predicted transmission window is set to 120 s when 𝑤 = 4, which is

the maximum duration observed with continuous data beacons with consistent link quality (§ 2.3.2.3).

The model generates 330 trees, each with a maximum depth of 5. We use the objective function

’multi:softprob’ to show each class’s probability distribution and the Mean Squared Error (MSE)

as the loss function. We design our model based on XGBoost [59] and SMOTE [60]. The inference

time per sample is 0.373 milliseconds on a Raspberry Pi 4 Model B [61]. Given the lightweight na-

ture of the model, it is particularly well-suited for deployment on resource-constrained devices. We

can anticipate exceptionally rapid inference times when the model is deployed at LoRa gateways.

29

SEMTECH SX1262 LoRa radioANT-916-CW RCL-SMA4.2 dBiAntennaSWARM RadioLoRa Radio3 km3 km2.5 Evaluation

Performance metrics: To evaluate the overall performance for uplink transmission, we focus on

the cumulative number of delivered packets as Throughput and the time from packet generation to

successfully delivered for each packet as Latency. For the link estimation model’s performance, we

rely on Accuracy and Mean Absolute Error (MAE) as metrics. We measure the performance of

individual gateway using Energy Eﬀiciency, represented as actual transmission attempts for each

packet. For multi-hop flooding protocol performance, we employ Latency as the key indicator to

present the packet delay among multiple gateways.

Baseline method: 1. COTS: We use the existing COTS protocol and network architecture of

Direct-to-satellite IoT devices as our baseline. Each gateway directly sends all the packets from

their satellite communication radio’s transmission buffer without any link estimation or link shar-

ing. This is evaluated and compared in energy efficiency evaluation (x2.5.2.2) and uplink data

transmission (x2.5.1). 2. LDB link estimation model: For link prediction, we only use the latest

data beacon (LDB) to guarantee agility at first. The input features include RSSI, SNR, ESP of

data beacon, noise, weather, satellite elevation, pass duration, and transmission window offset. 3.

ENV link estimation model: We build a linear regression model to predict the total delivered data

amount during one satellite pass duration with input from weather, satellite elevation, pass duration

and location index. We use the ENV model to compare energy efficiency performance (x2.5.2.2).

4. SateRIoT-𝑤: We use different link window lengths as our baselines. We select the 𝑤 (as men-

tioned in Section 2.3.2.2) values of 1,2,3,5 data beacon numbers to compare the different SateRIoT

scheme.

Default settings: The default packet generation frequency of each gateway is 1 packet per minute.

The packet includes multiple sensory data frames from its covered area. The default weather value

is 2, partly cloudy.

2.5.1 Overall Performance

In this section, we evaluate the overall performance in throughput and latency with SateRIoT

and COTS protocol. The delay and possible duplicate caused by link sharing are considered in

30

(a) Real Datasets Throughput

(b) Real Datasets Latency

Figure 2.11 Performance in real data of 4 gateways

calculating the latency and throughput according to the results from x2.5.3. Experiment A uses

real data collected simultaneously from 4 gateways in different outdoor locations. Experiment B

employs trace-driven datasets from 12 gateways to conduct emulation, involving 4 gateways trans-

mitting concurrently during 3 different time periods with similar satellite orbits and weather condi-

tions. Experiment C simulates 12 gateways, with ground-truth data generated by our accurate link

estimation model.

A. Real datasets experiments We use data collected outdoors over 100 minutes from 4 gate-

ways in x2.2.5 as ground-truths. We use our link estimation model and real collected information as

input to predict the link with real and emulate the throughput and latency. Figure 2.11a depicts the

throughput variations over time for all four gateways. As time progresses, the throughput disparity

between SateRIoT and the COTS method widens. After 100 minutes (6000 seconds), SateRIoT can

achieve 1.31(cid:2) cumulative sent data packet compared to COTS protocol. Additionally, Figure 2.11a

details the latency from packet generated time to delivery time. It is obvious that SateRIoT offers

substantially shorter latency compared to the COTS approach. Specifically, around 80% of the

packets can be transmitted in less than 8 minutes, compared to the COTS method, which requires

up to 24 minutes to achieve the same level of packet transmission. This results in SateRIoT being

3(cid:2) faster. All the packet in SateRIoT reaches a delay of less than 750 s while COTS extends to

3250 s. This indicates that the maximal latency of COTS can be up to 4.3(cid:2) longer than SateRIoT.

31

0100020003000400050006000Total Seconds050100150200250300350Cumulative Sent DataCOTSSateRIoT050010001500200025003000Delay (seconds)0.00.20.40.60.81.0CDFCOTSSateRIoT(a) Throughput of trace-driven data

(b) Latency of trace-driven data

Figure 2.12 Performance of trace-driven 12 gateways.

These findings affirm the superiority of SateRIoT in terms of throughput and latency in real-world

scenarios. The overall performance can be further improved when more gateways join the uplink

transmission.

B. Trace-driven experimental settings: To evaluate more gateways performance with SateRIoT,

We enlarge the gateways number to 12. We use 4 satellite radio devices as a group. Each deployed

3 times to emulate scenarios where gateways in 12 locations attempt to schedule data transmissions

during the same satellite pass. We select 3 satellite traces that were closely matched in terms of

elevation and pass duration. The maximum elevations and pass duration are 71(cid:14) with 31 minutes,

67(cid:14) with 29 minutes, and 69(cid:14) with 30 minutes, respectively. In each satellite trace, we deploy 4

gateways to collect real data sets from 4 distinct locations. The locations of the gateways vary

across the 3 satellite passes, ensuring coverage of 12 unique locations in total. This allows us to

effectively emulate a scenario where 12 gateways concurrently connect to the same satellite.

Results: Figure 2.12a shows that SateRIoT transmits 332 packets, which is 1.65(cid:2) more data than

the COTS baseline. Over time the data transmission gap between them widens because the base-

line focuses on individual transmissions, missing the opportunity to use the connection time fully.

SateRIoT efficiently schedules data and controls traffic, maximizing the use of the transmission

window across all gateways with a reliable connection. Figure 2.12b reveals that SateRIoT deliv-

ers all packets to the satellite within 357s, while the COTS protocol takes up to 1147s. Therefore,

32

02505007501000125015001750Total Seconds050100150200250300Cumulative Sent DataCOTSSateRIoT20040060080010001200Delay (seconds)0.00.20.40.60.81.0CDFCOTSSateRIoTFigure 2.13 Correlation heatmap of input features.

SateRIoT is up to 3.4(cid:2) faster, sending 80% of its packets in 280s compared to 700s for the COTS.

C. Generative datasets: Data availability for situations where elevation, pass duration, and weather

conditions are extremely similar is severely restricted. In order to broaden the scope of our emula-

tion experiments to encompass a wider range of satellite orbits and weather conditions, we deployed

2 or 4 satellite radios at 12 different locations, each tracking multiple satellite passes. This allows us

to generate datasets for simulating uplink data transmission among multiple satellite radios during

the same satellite pass. Figure 2.13 provides a clear overview of the interrelationships within the

original features directly recorded from collected real datasets. We can easily observe that RTrssi,

RTsnr, TDrssi, TDsnr exhibit strong positive correlations, while RTdelay and RTcount display notable

negative correlations. Additionally, pass duration, transmission window offset, and weather also

demonstrate noteworthy positive correlations with each other.

Based on the above observations, we establish a process to determine feature values using their

33

 5 7 B U V V L 5 7 B V Q U 7 ' B U V V L 7 ' B V Q U 1 R L V H 5 7 B G H O D \ 5 7 B F R X Q W ( O H Y D W L R Q ' X U D W L R Q 2 I I V H W : H D W K H U 5 7 B U V V L 5 7 B V Q U 7 ' B U V V L 7 ' B V Q U 1 R L V H 5 7 B G H O D \ 5 7 B F R X Q W ( O H Y D W L R Q ' X U D W L R Q 2 I I V H W : H D W K H U                                                                                                                                                                                                                            H                                                          H                                                                                                                                                                                                                                                                                                                                           (a) Throughput of 12 gateways with
max elevation 81(cid:14) in 45 min.

(b) Throughput of 12 gateways with
max elevation 66(cid:14) in 30 min.

(c) Throughput of 12 gateways with
max elevation 28(cid:14) in 14 min.

Figure 2.14 The throughput performance of generative data sets.

(a) Latency of 12 gateways with max
elevation 81(cid:14) in 45 min.

(b) Latency of 12 gateways with max
elevation 66(cid:14) in 30 min.

(c) Latency of 12 gateways with max
elevation 28(cid:14) in 14 min.

Figure 2.15 The latency performance of generative data sets.

inherent physical significance and a correlation heatmap. Initially, we randomly pick the maxi-

mum elevation from real datasets and then choose a pass duration that aligns with this elevation.

We randomly generate varying frequencies of connection time windows and set start times from

collected real datasets, considering the same maximum elevation and pass duration. Second, We

randomly select a value for RTsnr from the real datasets. Then we randomly choose a TDsnr from

the subset of real datasets that match the chosen RTsnr. Following this rule, based on determined

RTsnr and TDsnr, we randomly select RTrssi and TDrssi values. Third, we randomly chose RTdelay

from collected datasets, followed by a random RTcount based on consistent RTdelay. In addition, we

randomly set background noise RSSI from -106 dB to -87 dB ranging as real noise level.

Generative datasets experiments: We conduct further experiments to validate performance in

various satellite scenarios with the generative datasets. We set the maximum satellite elevations

to 81(cid:14), 66(cid:14) and 28(cid:14) with corresponding pass durations of 45, 30, and 14 minutes at 12 different

locations. Figure 2.14a, Figure 2.14b, and Figure 2.14c demonstrate that SateRIoT can achieve

34

05001000150020002500Total Seconds0100200300400500Cumulative Sent DataCOTSSateRIoT02505007501000125015001750Total Seconds050100150200250300Cumulative Sent DataCOTSSateRIoT0200400600800Total Seconds050100150Cumulative Sent DataCOTSSateRIoT0500100015002000Delay (seconds)0.00.20.40.60.81.0CDFCOTSSateRIoT0200400600800100012001400Delay (seconds)0.00.20.40.60.81.0CDFCOTSSateRIoT100200300400500600700800Delay (seconds)0.00.20.40.60.81.0CDFCOTSSateRIoTFigure 2.16 Throughput of large-scale simulation.

1.6(cid:2), 1.7(cid:2) and 1.9(cid:2) the data amount of baseline during one satellite pass. the data volume of the

baseline within a single satellite pass, respectively. Meanwhile, Figure 2.15a, Figure 2.15b, and

Figure 2.15c highlight that SateRIoT can achieve packet delivery latency that is 4.6(cid:2), 5.6(cid:2) and

3.93(cid:2) lower than the COTS protocol at most, respectively.

Large-scale simulation: We further simulate long-term application in 7 days. We select elevation

and pass duration from the real satellite orbit and emulate the performance of multiple satellite

passes using the generative datasets model. We repeat the 7-day experiment 100 times to obtain

large-scale simulation results to discover the performance gain of SateRIoT further. The boxplot

figures as shown in Figure 2.16. The accumulative packet volume of SateRIoT consistently exceeds

the COTS baseline over the 7 days, maintaining a steady range approximately from 1.5 to 2.0 (cid:2).

2.5.2 Link Estimation Model

In this section, we compare the performance among different link estimation trigger schemes

and different link estimation transmission window lengths. The ground truths for link estimation

during transmission windows are derived from local logs of real outdoor experiments. Then, we

use the link estimation model and input feature derived from the logs to predict transmission link

capability by comparing it with the real link. MAE is derived by comparing predicted values from

the link model to ground truths. We use w(windows) 1,2,3,4,5 to represent the data beacon amount

35

 ' D \   ' D \   ' D \   ' D \   ' D \   ' D \   ' D \   7 L P H                         3 D F N H W  $ P R X Q W 6 D W H / R 5 D % D V H O L Q HFigure 2.17 ACC and MAE of different link models.

during one transmission window.

2.5.2.1 Accuracy

We split our datasets randomly with 20% of the total as test datasets. The predicted accuracy

of window 4 can be up to 95.07% ACC with only 0.11 MAE. The feature importance of TD ESP

and RT ESP features are 0.16 and 0.13, respectively. This affirms the significance of our feature

engineering. The results affirm the reliability of our link estimation model, which can enhance the

performance of individual gateways and optimize overall data transmission.

Comparison among different models: The comparison results are shown in Figure 2.17.

the

accuracy is only 44.44% and 2.12 MAE for the LDB model. This suggests that the link information

from the data packet acknowledgement is necessary to guarantee accuracy. The ACCs of window

1,2,3,5 are 90.88%, 91.35%, 80.00%, 76.98% respectively. The MAEs of window 1,2,3,5 are 0.16,

0.23, 0.70 and 0.93 respectively. The performance for SateRIoT-1 and SateRIoT-2 are similar to

SateRIoT-4, whereas SateRIoT-3 performs poorly as it may interrupt continuous data pattern that

should have been captured in window 4, leading to inaccurate predictions of the last data beacon

with 18 classifications. When the window length exceeds 4, performance declines because only 2%

of transmission periods last longer than 120 seconds. This highlights the importance of selecting

an optimal window length to balance accuracy and efficiency. Based on empirical testing, we have

selected w=4 as our optimal window length.

36

 6 D 5 7   6 D 5 7   6 D 5 7   6 D W H 5 , R 7 6 D 5 7   / ' % 0 R G H O   6 D 5 7   U H S U H V H Q W  6 D W H 5 , R 7                 $ & &     $ & &     0 $ (             0 $ (Figure 2.18 ACC and MAE of different models with various elevation and locations

Performance with different maximum elevation and locations: To verify the performance in

different scenarios, we evaluated our model’s performance in 12 locations and multiple satellite

elevation values. Figure 2.18a displays the distribution of ACC and MAE for various groups of

maximum elevation angles, categorized into ”10-30”, ”30-50”, ”50-70”, and ”70-90” degrees. All

groups achieve more than 91% accuracy and a very low MAE of less than 0.3. This suggests that

our link estimation model is versatile and functions effectively across different elevation angles and

for satellites with varying orbits. Figure 2.18b presents the distribution of accuracy and MAE for

link estimation performance across 12 locations within a farm. Notably, our model achieves nearly

100% accuracy and zero MAE at 8 locations. Three locations surpass 87% accuracy, and one

location achieves 82%, with all MAEs remaining below 0.7. This deviation is within an acceptable

range and has minimal impact on the estimations and gateway performance. These results highlight

the robustness and reliability of our bursty link model across different elevations and geographic

locations.

Remark: Overall, the versatility of our model across varying elevation angles, combined with its

robustness in different geographic locations, demonstrates its efficacy in diverse operational con-

ditions. Balancing window length and triggering post-ACK receipt are key strategies to maximize

link estimation accuracy and system efficiency in LoRa networks. To improve adaptability and

scalability, it can be further trained in more complex environments.

2.5.2.2 Energy Eﬀiciency

Energy experimental settings: We conducted experiments using real datasets mentioned in x2.2.5

as test datasets. As outlined in Section 2.2.4, the COTS method attempts to transmit 6 times after

37

                     ( O H Y D W L R Q  $ Q J O H  5 D Q J H                          $ F F X U D F \          0 $ ( 6 D W H 5 , R 7    $ & & 6 D W H 5 , R 7    $ & & 6 D W H 5 , R 7  $ & & / ' %  $ & & 6 D W H 5 , R 7    0 $ ( 6 D W H 5 , R 7    0 $ ( 6 D W H 5 , R 7  0 $ ( / ' %  0 $ (                / R F D W L R Q                          $ F F X U D F \          0 $ (Figure 2.19 CDF comparison of energy efficiency and link waste.

the gateway receives each satellite data beacon. It continues attempting transmissions whenever a

satellite data beacon arrives. ENV stops the transmission if the cumulative packet amount reaches

its predicted value during the satellite pass duration. LDB predicts the data packet transmission

capability after one data beacon arrives and schedules predicted attempts. SateRIoT and SateRIoT-

n stop after successfully transmitting a predicted number of packets. Such an approach helps reduce

energy consumption by avoiding unreliable links. We predict and calculate the average attempts for

each packet as an energy efficiency metric during one satellite pass duration. Then, we count the

energy efficiency distributions from multiple satellite passes to compare the performance.

Results: Figure 2.19 compares the energy efficiency distributions between SateRIoT and the base-

line methods. SateRIoT consistently requires fewer than 1.65 attempts per packet across all scenar-

ios, in contrast to LDB, ENV, and COTS, which require up to 3.5, 6, and 6 attempts per packet,

respectively. For 80% of the cases, SateRIoT’s model prediction keeps energy efficiency under

1.25, while LDB and ENV, along with COTS, reach 1.6 and 1.85, respectively. Over 50% of the

cases with SateRIoT transmit packets without extra energy waste, whereas LDB achieves this in

32%. For each packet at one gateway, averagely, SateRIoT attains up to 3.3(cid:2) or 28.15 J lower en-

ergy usage than the COTS protocol during a single satellite pass. We can also observe that the model

with different window lengths exhibits a similar trend in energy efficiency, but SateRIoT performs

38

123456Energy (attempts/packet)0.00.20.40.60.81.0CDFCOTSLDBENVSateRIoT-1SateRIoT-2SateRIoT-3SateRIoT-5SateRIoT(a) One active gateway

(b) Two concurrent active gateways

(c) Three concurrent active gateways

Figure 2.20 The latency distribution of flooding beacons in multi-hop protocol.

slightly better than other SateRIoT-n solutions thanks to the higher accuracy. This demonstrates

that SateRIoT, unlike the more generalized link prediction methods like LDB or the COTS proto-

col, utilizes precise link estimation for energy-efficient transmissions. The overall energy saving

will increase as the number of gateways grows.

2.5.3 Multi-hop Flooding Protocol

Beacon flooding experiments: We conducted experiments to evaluate the performance of flooding

protocol in terms of reliability and latency. Initially, we designated one gateway in active mode and

initiated its beacon broadcast according to its unique transmission plan to all other gateway using

our CSMA-enabled multi-hop protocol. Then, we recorded the received time for all packets. This

process was repeated 10 times for each gateway, designating each one as the active gateway in turn.

For a setup of 12 gateways, we obtained latency data from a total of 30 experimental runs.

One active gateway results: Figure 2.20a depicts the overall latency distribution for all the LoRa

radios. The PDR is 100%. Approximately 70% of the standby gateways can receive the flooding

beacon from the active gateway within a latency of less than 2 s. Furthermore, around 94% of

the standby gateways can capture the flooding beacon from the active gateway within a 5-second

window. This indicates that in 94% of cases, beacon dissemination from two active gateways with

a time offset of larger than 5 s can prevent packet duplication. In our real data sets measured in

x 2.2.5, instances where a concurrent time offset of less than 5 s occur in less than 2% of all effective

transmission slots. This latency level is sufficiently low to facilitate timely data transmission.

Concurrent flooding beacons experiments: We seek to understand the performance when two

39

02468101214Beacon Delivery Delay (s)0.00.20.40.60.81.0CDFSingle Beacon0102030405060Queue Sync Delay (s)0.00.20.40.60.81.0CDFFirst BeaconSecond Beacon01020304050607080Queue Sync Delay (s)0.00.20.40.60.81.0CDFFirst BeaconSecond BeaconThird Beacongateways broadcast their beacons concurrently with a random time offset ranging from 0 s to 5 s.

We documented both the arrival time of the first beacon and the arrival times of both beacons at

each gateway. Moreover, to examine the extreme scenarios where three gateways become active

in close succession, we orchestrated an experiment where three beacons were broadcasted in rapid

succession with a random time offset ranging from 0 to 5 seconds, either between the first and

second beacon or between the second and third beacon. We record the arrival times of three beacons

for each gateway. Such instances accounted for less than 0.1% of our emulation data. This three-

beacon experiment was also repeated 120 times.

Latency Results: Figure 2.20b illustrates that approximately 90% of the gateways can receive both

beacons under 9 s, which may result in a maximum of 4 duplicate packets. Additionally, around

60% of the gateways can receive both beacons within 5 s. Figure 2.20c indicates that around 60%

beacons can be received with a delay of up to 20 s. around 60% beacons can be received with a

delay of up to 9 s. When comparing three figures in Figure 2.20, it’s evident that the time required

to reach approximately 80% of gateways approximately triples with the addition of each beacon.

Remark: The effective performance of the CSMA-enabled multi-hop flooding protocol enables

SateRIoT to operate efficiently. Its impressive time tolerance within 5 s ensures that concurrent

cases are infrequent and hardly impact overall data transmission or create duplicate issues.

Overall Energy Analysis: SateRIoT achieves a comparable throughput for each individual gate-

way while maintaining approximately half the energy consumption of the baseline method. The

propagation path between sensor nodes and satellites is usually hundreds of kilometers long, dra-

matically increasing the energy consumption to transmit the same amount of data. For example, the

power consumption of SWARM-M138 [34] modem is 12.24 J in Tx mode, while Semtech SX1262

LoRa radio [9] only consumes from 0.002 J to 0.045 J [62] with a 192-byte packet from SF7 to

SF12, which is 6120(cid:2) to 272(cid:2) less. In the busiest situations with the most power consumption

mode with SF12, one satellite transmits one packet; this requires data sharing with LoRa radio

transmission at most 12 times. The energy consumption of the LoRa packet is much lower than the

power consumption of satellite radio.

40

2.6 Related Work

Satellite Networking: L2D2 [25] presents a distributed scheduling system to reduce latency

for downlink data transmission with hybrid ground stations. Umbra [63] proposes a withholding

scheduling scheme for backhaul from satellites to large ground stations to overcome the uneven

queueing effect. Li et al. [64] evaluate that the orbit of COTS satellites is unpredictable because

of collision avoidance for Starlink [45]. However, SateRIoT does not heavily rely on orbit pa-

rameters but instead on coarse elevation and duration information. STARRYNET [65] builds an

open-source experimental framework to simulate complicated network behaviors. SpaceCore [26]

designs a stateless architecture to enable 5G deployment at satellites. Serval [66] enables near-real-

time insights for latency-sensitive imagery applications by emerging computational capabilities on

the satellites and ground stations. Compared to these works, SateRIoT focuses on the low-cost

satellite IoT in rural areas.

Satellite-based IoT: The research work on satellite IoT mainly focuses on network modeling and

physical layer design. Fraire et al. [67] present a sparse direct-to-satellite constellation design com-

bined with LoRa. Zhang et al. [68] utilize Bernoulli–Rician message to enable channel estimation

and user activity detection. Qian et al. [69] proposed a symmetry chirp spread modulation. Spec-

trumize [70] utilizes satellite movement-induced Doppler shift as a unique identifier for detection

and decoding. In contrast, SateRIoT designs and implements a network backhaul for rural area IoT

by space link.

2.7 Conclusion

To conclude, we introduce SateRIoT, a novel IoT backhaul architecture that merges LPWA tech-

nology on the ground with cost-effective IoT LEO satellites in space to support efficient rural area

networking. First, we design a bursty link model to predict packet transmission capacity, reducing

unnecessary data transmission. Next, we refine the model by selecting key features and optimizing

the window length. To achieve link sharing, we develop a multi-hop flooding protocol to main-

tain data sharing among all gateways and use a priority-based queue structure to avoid duplicate

transmissions. We implement SateRIoT with COTS satellite IoT and LoRa radios and evaluate the

41

performance on real deployment and real-world collected traces. The results show that SateRIoT

can achieve 3.3(cid:2) less energy consumption for an individual gateway. For overall performance,

SateRIoT reduces latency for packet delivery up to 5.6(cid:2) and improves overall throughput by 1.9(cid:2).

42

CHAPTER 3

DEMETER: RELIABLE CROSS-SOIL LPWAN WITH LOW-COST SIGNAL
POLARIZATION ALIGNMENT

3.1

Introduction

Internet-of-Things (IoT) plays a critical role in precision agriculture [71–77], in which soil

monitoring is an essential part. For example, understanding soil moisture levels in the crop root-

zone area can benefit irrigation water efficiency [78]; in-soil nitrate monitoring leads to a deeper

understanding of fertilization efficiency [75, 77, 79, 80]. Currently, commercial soil monitoring

systems [81–83] deploy underground soil sensors while leaving communication modules on the

ground. However, it is desirable to bury communication modules under the ground as well so

that they will not interfere with other agricultural activities (e.g., mowing, harvesting, fertilization,

irrigation) [84,85]. Such cross-soil communication has become a key feature of agricultural IoT [86,

87] and has attracted significant interests in recent years [72, 88–92].

Considering the large-range and long-term deployment in rural farms, rather than ad-hoc wire-

less sensor networks (WSN) [85, 91, 93, 94], Low-power Wide Area Networks (LPWAN) naturally

fit to enable agricultural IoT by embracing the wireless technologies featured by low energy con-

sumption and long communication distance [95]. LoRaWAN [21] is a popular LPWAN technology

operating on the unlicensed band. It supports long-range data collection from LoRa nodes (typi-

cally with sensors) to a gateway. Compared to other cellular LPWAN techniques (e.g., NBIoT [6],

LTE-M [96]), LoRaWAN provides a low-cost way to flexibly deploy new infrastructure in rural

farm areas where the cellular infrastructure is not available. Some measurement studies [88,97–99]

have shown the feasibility of adopting LoRaWAN to achieve cross-soil communication. However,

achieving a reliable cross-soil LoRaWAN in practice is not trivial. In our in-field measurement

(x 3.2.1), we observe that the soil condition changes (e.g., moisture, temperature) can lead to a sig-

nificant signal-to-noise ratio (SNR) variation up to 19.2 dB over a cross-soil LoRa link. Moreover,

according to our empirical study of the SNR variation in cross-soil LoRa communication (x 3.2.2),

compared to the variation of signal attenuation, polarization misalignment between the transmit-

43

Figure 3.1 The illustration of polarization misalignment problem (top) and the idea of cross-soil
communication and programmable polarization in Demeter (bottom).

ter and receiver antennas contributes more, leading to 13.2 dB SNR variation. For the available

configurations on commercial off-the-shelf (COTS) LoRa nodes [100], such an SNR variation will

shorten the communication distance by 4(cid:2) (x 3.5.2) and reduce energy consumption per day by up

to 82% (x 3.5.3).

The reason for cross-soil polarization misalignment is illustrated in Figure 3.1. LoRa nodes

and gateways adopt low-cost linearly polarized antennas. The polarization (i.e., blue arrows) of an

electromagnetic wave emitted from an antenna specifies the geometrical orientation of its oscil-

lations. The received signal power can be maximized if the polarization of the received signal is

aligned with the receiver’s antenna orientation (i.e., red arrows). However, the soil’s microscopic

structure and components keep changing over time, subject to different environmental conditions,

such as weather and agricultural activity. Consequently, as shown at the bottom of Figure 3.1,

when an electromagnetic wave travels through the soil with complicated scattering, refraction, and

reflection [101–105], its polarization changes over time. Even if the LoRa node’s antenna and

the gateway’s antenna are aligned in advance, the cross-soil transmission can cause polarization

misalignment between the received signal and the LoRa gateway’s antenna, degrading its power

44

ccPolarizationPlanexyzMisalignment.LoRanodeGatewayUndergroundAlignmentMisalignmentTxAntennaRxAntennaProgrammable Polarization.Transmitted SignalTraveling SignalRx OrientationFixed Polarization.Alignment.Propagation DirectionAgricultural ActivityWeatherAbovegroundTable 3.1 Qualitative comparison between Demeter and alternative methods regarding COTS com-
patibility (COTS.), cross-soil reliability, polarization self-alignment capability, and protocol over-
head in soil communication. “Pol.” and “Dir-Ante.” stand for “polarization” and “directional
antenna”, respectively.

Method

Challenge

COTS.

Dual Pol. [107]
Circularly Pol. [107]
Dir-Ante. [108, 109]
Scheduling [110, 111]
Demeter (Ours)

No
Yes
Yes
Yes
Yes

Cross-
soil
Reli-
ablity
High
Low
Low
High
High

Pol.
S-alig.
Capa-
ble
Yes
No
No
No
Yes

Protocol
Over-
head

Low
Low
Low
High
Low

significantly.

There is no prior work on cross-soil LoRa communications for the buried LoRa nodes to send

their sensor data to a LoRa gateway over a long distance. We want a solution that is low-cost,

interoperable with existing LoRa devices, reliable, and self-adjusting for polarization alignment,

with low overhead, such that the LoRa nodes buried underground can operate uninterruptedly for a

long time. For low cost and interoperability, we prefer a COTS-compatible solution. Specifically,

the LoRa technology is commercialized by Semtech [106] only, with its proprietary radio chips,

which all LoRa nodes and gateways in actual deployment use. We say a LoRa node (or gateway) is

COTS-compatible if it can be made from the Semtech LoRa radio chips with other (possibly custom-

designed) circuitry and computing/storage/signal processing components. A survey of relevant

technologies from the literature shows in Table 3.1 that none of them is satisfactory in meeting all the

challenges faced by cross-soil LoRa communications: COTS compatibility, reliability, polarization

self-alignment, and low overhead, which we elaborate below.

Challenge 1: COTS Compatibility. There is prior work using a dual-polarized antenna (consisting

of two orthogonal polarization antennas) and two RF chains to fully collect the energy of arbitrary

polarization signals [107]. Although the work was not directly on LoRa, it is conceivable that

its method could in principle be applied on a LoRa gateway to receive polarization-misaligned

signals from an underground LoRa node. However, this method is not COTS-compatible because

all existing COTS LoRa gateways [112–114] only have one RF chain in their radio chips, not two

45

RF chains required in [107]. Hence, not only does this method carry a higher cost, but it is not

applicable to the existing deployments using COTS gateways.

Challenge 2: Cross-soil Reliability. In aboveground and open-space communications, LoRa nodes

can utilize circularly polarized antennas, which generate circular polarization signals, to mitigate

the polarization misalignment problem [107, 115]. However, the signal reflection, refraction, and

scattering of under-ground communications are much more complex, leading to dramatic axial ratio

distortion of circular polarization [116–119]. This distortion means the wave is no longer circular

polarization and can result in a severe signal loss at LoRa gateways, degrading communication

reliability [118–123].

Challenge 3: Polarization Self-alignment Capability. Low-cost electronically switchable direc-

tional (ESD) antennas [108, 109, 124, 125], providing antenna gain in a certain direction, can be

equipped on LoRa nodes to enhance SNR. However, these directional antennas carry a higher cost

in the LoRa band, and they are not capable of mitigating polarization misalignment. Moreover,

the antenna gain from low-cost ESD antennas in the low-frequency band is usually too limited to

tolerate the observed 13.2 dB SNR variation completely [126–130].

Challenge 4: Protocol Overhead. LoRa nodes can estimate cross-soil links and schedule trans-

missions under favorable conditions for good polarization [110, 111]. However, considering the

continuously changing soil conditions, maintaining an accurate estimation and scheduling the trans-

missions may bring considerable computation and energy overhead, particularly for low-cost LoRa

nodes.

To address the above challenges, in this paper, we propose Demeter, a low-cost and low-power

polarization-programmable antenna system on COTS-compatible LoRa nodes to enable reliable

cross-soil LoRa communication. It is inter-operatable with the LoRa gateways in existing deploy-

ment. Our key idea is to design a single RF-chain compatible circuit to dynamically adjust the initial

polarization of LoRa node transmission. As shown at the bottom of Figure 3.1, considering the po-

larization changes during underground signal propagation, the LoRa node adjusts the polarization

of its transmitted signal dynamically to ensure the polarization of the received signal is aligned with

46

LoRa gateway’s antenna.

The design of Demeter involves three key problems.

First, due to the single RF chain resource on existing LoRa nodes, it is not trivial to design a

circuit that enables flexible polarization adjustment. To address this challenge, Demeter uses an

RF splitter to split the raw signal from the single RF chain into two signals. In addition, Demeter

adopts a dual-polarized antenna, which takes the two signals as the inputs of its two orthogonal

polarization dipole antennas to generate a linearly polarized transmission signal. The polarization

of the transmission signal can be adjusted by configuring the amplitude ratio of the two signals.

Second, considering the energy constraints on COTS LoRa nodes, energy-exhausting hardware

components (e.g., voltage attenuator, amplifier) should not be involved in signal amplitude control.

How to control the amplitudes of the two split signals with low-power hardware components is an-

other challenge. To address this challenge, Demeter converts an amplitude ratio to a configurable

phase offset between the two signals by utilizing a low-power digital phase shifter and a passive

hybrid coupler. The phase shifter adds a pre-configured offset to the phase of one signal. More-

over, Demeter uses the hybrid coupler to mix the two signals to generate two new signals with the

corresponding amplitude ratio for the two orthogonal dipole antennas and further customize the

hybrid coupler for the LoRa frequency band.

Third, due to the occasional change of the soil condition, the polarization of cross-soil signals

is changing over time. To lower link maintenance overhead while keeping agile to the polarization

change, when to trigger and how to calibrate the end-to-end polarization alignment between LoRa

nodes and gateways are not trivial. To address these challenges, Demeter triggers a calibration

process adaptively according to potential soil condition changes in a day. A LoRa gateway initializes

a polarization calibration process and asks the targeted LoRa node to transmit multiple beacons

with different polarizations. Then we develop a heuristic algorithm to search the best polarization,

balancing the searching accuracy and energy overhead on LoRa nodes.

We implement a Demeter prototype with a COTS digital phase shifter [131], a COTS RF split-

ter [132], a COTS dual-polarized antenna [133], and a customized hybrid coupler on a COTS LoRa

47

node with Semtech SX1276 radio [134]. We have conducted extensive experiments in various soil

environments. Our results show that compared with standard LoRa, Demeter can achieve up to

9.94 dB SNR gain outdoors, 4(cid:2) horizontal communication distance, and at least 20 cm under-

ground buried depth improvement.

In summary, our contributions are summarized as follows:

• To the best of our knowledge, we are the first to develop a reliable cross-soil LPWAN, which

is an essential part of agricultural IoT and in-situ soil sensing for realizing precise agriculture

in scale.

• We propose a low-cost and low-power polarization-programmable antenna system on COTS

single RF chain LoRa nodes. Moreover, we design practical methods to monitor and calibrate

polarization alignment over a cross-soil LoRa link with low maintenance costs.

• We implement the prototype of Demeter and evaluate it in real-world deployments. The

results show an SNR gain of up to 9.94 dB in real environments, 4(cid:2) communication distance,

and at least 20 cm deeper deployment depth improvement on average. Moreover, Demeter

decreases energy consumption per day by up to 82%.

3.2 Preliminary and Motivation

LoRa is a wireless communication technology that uses chirp spread spectrum (CSS) modula-

tion to overcome the effects of narrowband interference and multipath fading [135]. This allows

LoRa devices to communicate over long distances with low power consumption [95, 136]. Fig-

ure 3.2 illustrates a typical cross-soil LoRa communication system, where LoRa nodes are buried

in underground soil and LoRa gateways are deployed in aboveground air. The sensors on the LoRa

nodes generate hourly/daily sensory data that is sent to the gateways over a cross-soil link. Many

factors (e.g., rainfall, snowfall, temperature, animal activity, plant growth) incur significant soil

condition changes over time, leading to unstable cross-soil LoRa links and resulting in packet loss

and energy waste.

48

Figure 3.2 The illustration of cross-soil LoRa communication system.

3.2.1

In-field Cross-soil Link Measurement

Agriculture soils are categorized into more than ten types, as determined by the ratio of three

mineral components (i.e., sand, silt, and clay) according to USDA soil texture document [137,138].

For example, clay soil consists of more than 50% clay while loam has about 40% sand, 40% silt,

and 20% clay. In agriculture, loam soil holds a significant position due to its ideal balance of sand,

silt, and clay particles. This balance affords loam soil excellent properties for plant growth, making

it highly desirable for farming and gardening.

We conducted a 9-day in-field experiment to explore the quality of cross-soil LoRa communi-

cation in a wheat farm. The soil is CvraaB-Conover loam. Its mineral composition is about 40%

sand, 40% silt, and 20% clay, respectively [137–139]. Rainfall comes on Day 2 and Day 9. A

moisture sensor [140] connected to a COTS LoRa node is buried at 30cm depth underground, and

we deploy a COTS LoRa gateway aboveground. The horizontal distance between the LoRa node

and gateway is 10m. The LoRa node sends a soil moisture reading every 30 minutes. The gateway

extracts RSSI (Received Signal Strength Indicator), SNR, and VWC (Volumetric Water Content)

that measure the ratio of the volume of water to the unit volume of soil from each received packet.

Figure 3.3 shows the results. From VWC curve, we can observe two rainfalls (i.e., VWC shift-

ing up) on Day 2 and Day 9. Correspondingly, the SNR and RSSI decrease sharply on Day 2 and

fluctuate on Day 9. Moreover, we observe another RSSI and SNR drop and fluctuation on Day 6

and Day 4, respectively. Based on the historical weather record, abrupt temperature changes appear

49

AbovegroundUndergroundLoRa NodeRoot growthMicroorganismMoistureRain FallSnow FallAnimals ActivityPlant GrowthGatewayFigure 3.3 9-days performance of in-field cross-soil communication.

these days. Overall, the SNR and RSSI are extremely unstable across 9 days. The RSSI difference

reaches to 21.32 dBm. The SNR ranges from 10.4 dB to -8.8 dB. The total 19.2 dB SNR varia-

tion verifies that soil condition (e.g., moisture, temperature) changes can degrade communication

reliability in real-world deployments.

3.2.2 Cross-soil Signal Polarization Study

For the cross-soil LoRa link in our in-flied measurement, the observed dramatic SNR variation

consists of two parts. One part is from the change of signal attenuation determined by soil’s di-

electric properties [94, 141, 142] that keep changing along with the changes of soil condition (e.g.,

moisture, temperature). The other part is from signal polarization misalignment [115]. COTS LoRa

nodes and gateways adopt linearly polarized antennas. The polarization of an electromagnetic wave

describes the orientation of the oscillating electric field. The electromagnetic wave of LoRa signals

is transverse and travels in a specific direction. A magnetic field and an electric field are perpendic-

ular to each other and the direction of propagation. The linear polarization indicates the orientation

of the electric field, which oscillates in a fixed plane determined by the magnetic and electric fields

and can be purely vertical, horizontal, or any angle in between. When the polarization of the re-

ceived signals is well aligned with a LoRa gateway’s antenna orientation, the energy of the electric

fields can be fully captured by the receiving antenna. Consequently, the SNR of the received sig-

nals can be maximized. However, in cross-soil communication, signal scattering, reflection, and

refraction are inevitable during signal propagation in soils, making the polarization variation at the

50

-110-100RSSI(dBm)-12-60612SNR(dB)Day1Day2Day3Day4Day5Day6Day7Day8Day9Time0102030VWC(cm3/cm3)Figure 3.4 SNR variation analysis across multiple VWC settings.

receiving antenna [104, 143–146], resulting inevitable polarization misalignment.

We have conducted controlled experiments to investigate the SNR variation incurred by polar-

ization misalignment in cross-soil communication. In an indoor environment, we bury a COTS

LoRa node in a plastic container of sandy soil (i.e., > 90% sand) and utilize a COTS LoRa gateway

to receive packets. To imitate the soil moisture changes in our in-flied measurement, we change

VWC from 0% to 48%. When VWC is 48%, the soil tends to be water-saturated, which only hap-

pens when flooding appears in the real world. Thus, we do not set a higher VWC. To verify the

influence of polarization misalignment, for each VWC, we manually vary the polarization of the

received signals by uniformly rotating the antenna of the LoRa gateway to 16 different orientations

in a plane perpendicular to the direction of electromagnetic wave propagation. For each antenna

orientation, the LoRa node will transmit 10 packets. Then, the LoRa gateway collects the packet

SNR.

In Figure 3.4, the five-value boxes show the observed SNR distribution at different VWC con-

figurations. We can see that the SNR variation between the maximum and the minimum SNR under

a VWC configuration is 13.2 dB on average. Such a huge SNR variance is purely caused by po-

larization misalignment. Moreover, for each VWC configuration, the maximum SNR is achieved

when the polarization is well aligned. Thus, the changes in the maximum SNR reflect the changes

in signal attenuation. We can see the maximum SNR monotonously decreases from 7.0 dB to -

51

0%8%16%24%32%40%48%VWC (cm3/cm3)-13-10-7-4-1258SNR (dB)OverallPos1Pos2Pos35.6 dB, namely the signal attenuation increases 12.6 dB, when VWC increases from 0% to 48%.

In our in-flied measurement, the VWC is in the range of [7.8%, 23.1%]. The SNR variation caused

by signal attenuation is about 5.8 dB which is 2.3(cid:2) less than the 13.2 dB SNR variation brought by

polarization misalignment. By adding both variations, the total SNR variation is 19.0 dB, which is

relatively consistent with the observed 19.2 dB.

Furthermore, we verify that the signal polarization is indeed changed by soil condition change

(i.e., VWC increase). The red, yellow, and purple dashed lines represent the SNR of three specific

antenna orientations.

If the cross-soil signal polarization is constant under different VWC, the

trend of the three dashed lines should be similar to the monotonous decrease of the maximum SNR.

However, the patterns of the three dashed lines are non-monotonic and totally different. Specifically,

the standard deviations of the SNR difference between each dashed line and the maximum SNR are

4.66 dB, 4.2 dB, and 3.78 dB. The dynamic pattern and high deviation verify that the soil condition

changes dramatically influence the quality of received signals.

3.2.3 Motivation

The observed reliability vulnerability motivates us to compensate for the polarization misalign-

ment in cross-soil LoRa communication. On the other hand, a LoRa gateway may cover hundreds

or thousands of LoRa nodes. Different LoRa links have diverse polarization misalignments. If

we compensate for the polarization misalignment at the gateway side (e.g., rotating the gateway’s

antenna), the gateway has to assign different LoRa nodes’ transmissions at different times. The cal-

ibration process, transmission scheduling, and network synchronization will consume considerable

energy on LoRa nodes, compromising the energy benefits of the achieved reliable data transmission

and incurring new scalability concerns. This motivates us to compensate for polarization misalign-

ment at the LoRa node side compatible with the standard LoRaWAN media access control (MAC).

3.3 System Design

Figure 3.5 provides an overview of Demeter. Demeter consists of three parts: 1) the hardware

design of a programmable antenna that allows the LoRa node to configure its signal polarization;

2) a polarization alignment calibration algorithm that generates the optimal LoRa node’s setting to

52

Figure 3.5 System Overview of Demeter.

achieve polarization alignment at the LoRa gateway; and 3) a link calibration scheduling method

that optimizes system energy efficiency while keeping communication reliability.

3.3.1 Hardware Architecture

Figure 3.6 illustrates Demeter’s hardware architecture, which consists of four components as

follows:

Passive RF Splitter: To achieve an adjustable polarization degree with only one RF chain and

dual-polarized antenna, the first step is to split the single-channel signal into two channels. We use

a 3 dB wide-band 2-way passive splitter to divide a signal into two identical signals, each with half

the amplitude of the raw signal.

Phase Shifter: An adjustable phase shifter adds a phase offset to the input signal. Demeter adopts

a programmable digital phase shifter [131] which can be embedded into COTS LoRa nodes with a

3.3 V voltage supply.

Hybrid bridge Coupler: The hybrid coupler is implemented as a four-ports (i.e., two inputs, two

outputs) 3 dB 90(cid:14) bridge with center frequency at the US915 band for LoRa. A hybrid coupler

divides the signal of each input port into two signals with the same amplitude and 90(cid:14) phase dif-

ference, then crossly combines two signals from two different input ports to generate the signals at

two output ports.

Dual-polarized Antenna: A dual-polarized antenna consists of two dipole antennas with orthog-

onal linear polarization orientations and an overlapping geometrical center. The simple structure

makes it easy to manufacture and has been widely used in wireless communication.

53

LoRanodeLink Calibration SchedulingPolarization Alignment CalibrationGatewayProgrammable AntennaCross-soilPropagationFigure 3.6 Demeter’s hardware architecture and adjustable polarization principle.

3.3.2 Adjustable Polarization Principle

As shown in Figure 3.6, the raw signal transmitted by a LoRa node is 𝑋0 with amplitude 𝐴 and

phase 𝛼, which can be expressed as 𝑋 = 𝑓 „ 𝐴, 𝛼, 𝑡”, where 𝑡 is the time. First, we use the passive RF
splitter to divide the raw LoRa signal into two identical signals with half amplitude: 𝑋1 = 𝑓 „ 𝐴
2
and 𝑋2 = 𝑓 „ 𝐴
2
offset 𝜙. We assume the phase shifter is ideal without extra loss, the amplitude of 𝑋1 is still 𝐴
we have 𝑋0

, 𝛼, 𝑡”. For the first way signal 𝑋1, the digital phase shifter is used to add a phase

, 𝛼 ‚ 𝜙, 𝑡”. We can see that the phase difference between 𝑋0

2 . Then

, 𝛼, 𝑡”

1 and 𝑋2 is 𝜙, too.

1 = 𝑓 „ 𝐴

2

The port1 and port2 of the hybrid coupler take 𝑋0

1 and 𝑋2 as input, respectively. The output
signals at port3 and port4 are 𝑋3 and 𝑋4. To calculate the value of 𝑋3 and 𝑋4, we use 𝑋 𝑗𝑖 to

represent the partial output signal at port 𝑖 from the input signal at port 𝑗. As shown in Figure 3.6,

𝑋3 consists of two parts: 𝑋32 (i.e., the blue arrow at port3) and 𝑋31 (i.e., the red arrow at port3).
The hybrid coupler adds a 90(cid:14) phase offset on 𝑋2 to generate 𝑋32 (i.e., the blue crossing line). Then,
𝑋3 can be calculated as follows:

ﬁ𝑋3 = ﬁ𝑋32 ‚ ﬁ𝑋31 = 𝑓 „

𝜋

𝐴

, 𝛼 ‚

, 𝑡” ‚ 𝑓 „

, 𝛼 ‚ 𝜙, 𝑡”

2
Port3 vector ﬁ𝑋3 (i.e., the brown arrow in Figure 3.6) is vectors sum of ﬁ𝑋32 and ﬁ𝑋31.

ﬁ𝑋32, ﬁ𝑋31
compose a rhombus with same amplitude. Therefore, the phase value of ﬁ𝑋3 should be the mean

2

2

(3.1)

(3.2)

(3.3)

, 𝛼 ‚

, 𝑡”

𝜋

2

, 𝛼 ‚ 𝜙, 𝑡”

𝐴

2
𝐴

2

ﬁ𝑋32 = 𝑓 „

ﬁ𝑋31 = 𝑓 „
𝐴

54

X!132490°Hybrid CouplerX"Passive RF SplitterΦ(0°-180°)Dipole ADipole ADipole BDipole BX#-3dB-3dBX"′X$X%ΦΦ + 90°θΦθA3A4ϒPort1Port2Port3Port4Polarization ComposePolarization PlanePropagation Direction90°yxzPhase Shifterphase of ﬁ𝑋32 and ﬁ𝑋31. Let the amplitude of ﬁ𝑋3 be A3. We can simplify ﬁ𝑋3 as:

ﬁ𝑋3 = 𝑓 „ 𝐴3, 𝛼 ‚

𝜙

2

‚

𝜋

4

, 𝑡”

(3.4)

Similarly, 𝑋4 is the combination of 𝑋42 (i.e., the blue arrow at port 4) and 𝑋41 (i.e., the red arrow
1, the hybrid coupler adds a 90(cid:14) phase offset to generate 𝑋41 (i.e., the red

at port 4). Based on 𝑋0

crossing line). We have 𝑋4 as follows:

ﬁ𝑋41 = 𝑓 „

𝐴

2

, 𝛼 ‚ 𝜙 ‚

, 𝑡”

𝜋

2

𝐴

2

, 𝛼, 𝑡”

ﬁ𝑋42 = 𝑓 „
𝐴

2

ﬁ𝑋4 = 𝑋41 ‚ 𝑋42 = 𝑓 „

, 𝛼 ‚ 𝜙 ‚

, 𝑡” ‚ 𝑓 „

𝜋

2

𝐴

2

, 𝛼, 𝑡”

Let us indicate the amplitude of ﬁ𝑋4 as A4,

ﬁ𝑋4 = 𝑓 „ 𝐴4, 𝛼 ‚

𝜙

2

‚

𝜋

4

, 𝑡”

(3.5)

(3.6)

(3.7)

(3.8)

We can see that 𝑋3 and 𝑋4 have the same phase while their amplitudes A3 and A4 are different.

We further feed 𝑋3 and 𝑋4 to the two feeding ports of the two dipole antennas of the dual-

polarized antenna. Then the two dipole antennas produce two orthogonal electric fields. Generally,

an electromagnetic plane wave that moves in the 𝑧-axis direction has two components in 𝑥-axis and

𝑦-axis [147].

𝐸𝑥 = 𝐸𝑥𝑚𝑐𝑜𝑠„𝑤𝑡 (cid:0) 𝑘 𝑧 ‚ 𝛽𝑥”

𝐸𝑦 = 𝐸𝑦𝑚𝑐𝑜𝑠„𝑤𝑡 (cid:0) 𝑘 𝑧 ‚ 𝛽𝑦”

(3.9)

(3.10)

ﬁ𝐸 = ﬁ𝑒𝑥 𝐸𝑥 ‚ ﬁ𝑒𝑦𝐸𝑦 represents for electric field and 𝑚 means magnitude/amplitude. At the two dipole

antennas in the dual-polarized antenna of Demeter, we have two orthogonal electric fields in the
𝑥-𝑦 plane. Both the wave phases in 𝑥 and 𝑦 are equal to 𝛽 = 𝛼 ‚ 𝜙
2

4 . We set the same 𝑧 to 0 here.

‚ 𝜋

√

√

jj ﬁ𝐸 jj =

𝐸 2

𝑥 ‚ 𝐸 2

𝑦 =

𝐸 2

𝑥𝑚 ‚ 𝐸 2

𝑦𝑚𝑐𝑜𝑠„𝑤𝑡 ‚ 𝛽”

𝛾 = 𝑎𝑟𝑐𝑡𝑎𝑛„

𝐸𝑦
𝐸𝑥

” = 𝑎𝑟𝑐𝑡𝑎𝑛„

𝐴3
𝐴4

”

55

(3.11)

(3.12)

According to the principle of vector composition, we can change the amplitude ratio 𝐴3

𝐴4 between
𝑋3 and 𝑋4 by adjusting the phase offset 𝜙 to determine the degree of linear polarization 𝛾 (i.e.,

the green arrow at polarization compose). The propagation direction (i.e., black dashed arrow) of

electromagnetic waves is the normal of the polarization plane. Finally, we can generate signals with

different linear polarization degrees.

3.3.3 Coupler Architecture

Some hardware components of our system, such as the phase shifter, the passive RF splitter,

and the dual-polarized antenna are COTS devices that can be obtained easily. The hybrid coupler

for a LoRa node should be low-cost with a small size without any power supply. However, the sub-1

GHz bridge coupler which meets our demand is limited on the market. Thus, we design the bridge

coupler by ourselves.

The purpose of the coupler is to produce 90(cid:14) phase difference between the signals of two output

ports when the input signal is only from one input port and the other one is isolated. Three kinds

of RF bridge couplers can achieve the target: codirectional, contradirectional, and transdirectional

with different isolated port positions relative to input ports. Compared to the other two solutions,

as illustrated in Figure 3.7a, a transdirectional coupler can avoid crossover and direct current bias

problems because the isolated port and input port are on the same side, making the coupler smaller

and scalable for IoT devices [148].

Transdirectional Coupled Line. We adopt an even-odd mode analysis [148] at four ports of the

hybrid coupler in Figure 3.7. Port1 is input and port2 is isolated while port3 is the through and

port4 is coupled. 𝑉𝑖 (i=1,2,3,4) represents the voltage at port𝑖 and 𝑉0 is the generator voltage. Let
𝑖𝑛 be the input impedance at port1 for the even mode and 𝑍 𝑜
𝑍 𝑒
mode. 𝑍0 is the load impedance of the transmission line in this coupler.

𝑖𝑛 be the input impedance of the odd

𝑍 „𝑒,𝑜”
𝑖𝑛

= 𝑍0„𝑒,𝑜”

𝑍0 ‚ 𝑗 𝑍0„𝑒,𝑜”𝑡𝑎𝑛𝜃
𝑍0„𝑒,𝑜” ‚ 𝑗 𝑍0𝑡𝑎𝑛𝜃

𝑉 „𝑒,𝑜”
1

= 𝑉0

𝑍 „𝑒,𝑜”
𝑖𝑛
𝑍 „𝑒,𝑜”
𝑖𝑛

‚ 𝑍0

56

(3.13)

(3.14)

(a) Hybrid coupler geometric structure

(b) Equivalent circuit for each unit cell

Figure 3.7 The illustration of the hybrid coupler design.

According to the symmetrical structure of the transdirectional coupler, the voltage at port4 is

𝑉4 = 𝑉 𝑒
4

Based on the definition of 𝑍𝑖𝑛

‚ 𝑉 𝑜

4 = 𝑉 𝑒

1

(cid:0) 𝑉 𝑜

1 = 𝑉0 »

𝑍 𝑒
𝑖𝑛
𝑍 𝑒
𝑖𝑛 ‚ 𝑍0

(cid:0)

𝑍 𝑜
𝑖𝑛
𝑍 𝑜
𝑖𝑛 ‚ 𝑍𝑜

…

𝑍𝑖𝑛 =

𝑉1
𝐼1

=

√

𝑍0 =

𝑍0𝑒 𝑍0𝑜 =

‚ 𝑉 𝑜
𝑉 𝑒
1
1
‚ 𝐼 𝑜
𝐼 𝑒
1
1
√
𝑖𝑛𝑍 𝑜
𝑍 𝑒
𝑖𝑛 = 𝑍𝑖𝑛

Combined with Equation 3.15. 𝑁 is the coupling coefficient.

𝑉4 = 𝑉0

𝑗 „𝑍0𝑒 (cid:0) 𝑍0𝑜”𝑡𝑎𝑛𝜃
2𝑍0 ‚ 𝑗 „𝑍0𝑒 ‚ 𝑍0𝑜”𝑡𝑎𝑛𝜃
𝑍0𝑒 (cid:0) 𝑍0𝑜
𝑍0𝑒 ‚ 𝑍0𝑜
𝑗 𝑁

𝑁 =

1 (cid:0) 𝑁 2𝑐𝑜𝑠𝜃 ‚ 𝑗 𝑠𝑖𝑛𝜃

𝑉4 = 𝑉0

p

Similarly, we calculate 𝑉3 for port 3 as follows:

𝑉3 = 𝑉 𝑒
3

‚ 𝑉 𝑜

3 = 𝑉0

p

p

1 (cid:0) 𝑁 2
1 (cid:0) 𝑁 2𝑐𝑜𝑠𝜃 ‚ 𝑗 𝑠𝑖𝑛𝜃

57

(3.15)

(3.16)

(3.17)

(3.18)

(3.19)

(3.20)

(3.21)

1 Input2Isolation4Coupled3 Throughθ′𝟐𝐂𝐒𝐂𝒈θ′𝟐𝐂𝒈𝐙′𝟎𝐞𝟐𝐂𝐒𝐂𝒆𝐙′𝟎𝐨𝐂𝒐Even ModeOdd ModeWe can notice that there is always 90(cid:14) phase difference between 𝑉3 and 𝑉4 due to the same denom-
inator and orthogonal numerator. The scattering parameter 𝑆 can be computed with voltage ratios

(𝑆𝑖 𝑗 =

𝑉 𝑗
𝑉𝑖 ).

𝑆 =















p

(cid:0)

0

0

1 (cid:0) 𝑁 2 𝑗

𝑁

0

0

𝑁

p

(cid:0)

1 (cid:0) 𝑁 2 𝑗

p

1 (cid:0) 𝑁 2 𝑗

(cid:0)

𝑁

0

0

𝑁

p

1 (cid:0) 𝑁 2 𝑗

(cid:0)

0

0















(3.22)

Then, to keep the absolute value of 𝑆31, 𝑆32, 𝑆41, and 𝑆42 identical for generating the designed 𝑋3

and 𝑋4 signals, we set 𝑁 as

2 . The 𝑆 parameter of our coupler becomes:

p
2

𝑆 =

0 (cid:0) 𝑗

1

0

0









(cid:0) 𝑗






0

1

1 (cid:0) 𝑗















(3.23)

1 (cid:0) 𝑗

0

0

0

0

Periodically Loaded Design. For a conventional coupled line, we need to satisfy Equation 3.17

under the necessary constraint of the transdirectional operation:

𝜃𝑒 (cid:0) 𝜃𝑜 = „2𝑛 ‚ 1”𝜋

(3.24)

However, it is challenging to adjust 𝜃𝑒, 𝜃𝑜, 𝑍0𝑒 and 𝑍0𝑜 simultaneously because the spacing between

the coupled lines would impact both the impedances and phase velocities. Inspired by transdirec-

tional coupled-line couplers working on 3.6 GHz [149], we adopt periodically loaded coupled lines

to overcome this challenge.

We design multiple unit cells with capacitors between coupled lines. The unit cell, consisting of

several capacitors and transmission lines, is shown in Figure 3.7a. The orange rectangles are patch

capacitors. Here 𝑁 is the number of unit cells. Let 𝜃0 be the electrical length of each unit cell and

𝑍0
0 the characteristic impedance of the transmission line as the equivalent circuit that Figure 3.7b
shows. 𝐶 is the capacitance between the coupler transmission lines. 𝐶𝑔 is the capacitance between

the coupler and the ground. The shunt susceptances of periodic load on even and odd modes are 𝑙𝑒

58

(a) Rogers5880

(b) FR4

Figure 3.8 Coupler efficiency with different substrates.

and 𝑙𝑜. According to the design in [149], we have:

𝑐𝑜𝑠

𝜃
𝑁 = 𝑐𝑜𝑠𝜃0 (cid:0)
√

𝑏

𝑠𝑖𝑛𝜃0

2
2𝑠𝑖𝑛𝜃0 ‚ 𝑏𝑐𝑜𝑠𝜃0 (cid:6) 𝑙
2𝑠𝑖𝑛𝜃0 ‚ 𝑏𝑐𝑜𝑠𝜃0 (cid:7) 𝑙

𝑍0 = 𝑍0
0

0„𝑜,𝑒”

𝑙„𝑜,𝑒” = Ω𝐶𝑜,𝑒 𝑍0
𝐶𝑜 (cid:0) 𝐶𝑒
2

𝐶𝑠 =

(3.25)

(3.26)

(3.27)

(3.28)

According to Equations 3.25, 3.26, 3.27, and 3.28, the electrical length for each unit cell can be

assigned with a preconfigured value. Then we can calculate the capacitance value of 𝐶𝑠 accordingly

for the LoRa frequency band.

HFSS Simulation: We use the High-Frequency Structure Simulator (HFSS) to perform electro-

magnetic simulation on our system. Initially, we use Roger5880 0.5mm thickness substrate with

low signal energy loss to ensure high efficiency and verify the system. Figure 3.8a shows the effi-

ciency of the coupler across the four ports in the US915 band. 𝑆31 and 𝑆41 can achieve -3.04 dB

and -3.24 dB, which are consistent with half amplitude split of 𝑋31 and 𝑋41. And the reflection loss

in 𝑆11 and 𝑆21 are ultra-low as -28.13 dB and -26.52 dB, achieving high energy efficiency. How-

ever, Rogers5880 is expensive hindering us from deploying Demeter on a large scale. In addition,

Rogers5880 is too soft to be distorted easily, causing unexpected circuit changes and high loss.

59

0.70.80.911.1Frequency(GHz)-42-35-28-21-14-70Efficiency(dB)dB(S11)dB(S21)dB(S31)dB(S41)0.70.80.911.1Frequency(GHz)-42-35-28-21-14-70Efficiency(dB)dB(S11)dB(S21)dB(S31)dB(S41)(a) Phase offset for port1

(b) Phase offset for port2

Figure 3.9 Cross phase shift in Demeter’s coupler.

To benefit IoT deployment, we redesign our coupler with substrate FR4 at 1.6mm thickness,

which is one of the cheap, hard, and easily processed substrate materials. In HFSS, by adjusting the

geometrical shape and parameters like electrical length, bridge width, and capacitors, we success-

fully achieve a 90(cid:14) phase offset with low loss as shown in Figure 3.9. The phase difference between
𝑆32 and 𝑆42 range from to 89.85(cid:14) to 90.22(cid:14) in the US915 band and 𝑆32 and 𝑆42 range from to 90.6(cid:14)
to 90.92(cid:14). Figure 3.8b shows that 𝑆31 and 𝑆41 are -3.35 dB and -3.24 dB. And the reflection loss
in 𝑆11 and 𝑆21 are -22.23 dB and -22.29 dB, still very low levels. The performance is comparable

with Roger5880.

3.3.4 Polarization Alignment Calibration

The calibration problem is to find the optimal phase offset configuration of the phase shifter

that makes the signal polarization aligned between a Demeter node and a LoRa gateway, namely

the observed SNR is the highest at the LoRa gateway.

Instead of traversing all possible phase

offset configurations, we design a heuristic SNR peak searching algorithm to minimize the control

overhead.

SNR Peak Searching Problem: The digital phase shifter can provide 𝑛 different discrete phase

offsets, which generate different polarization degrees of the output signal. We use the term “phase

index” (i.e., from 1 to 𝑛) to represent these phase offsets in ascending order. We aim to find out the

60

0.80.850.90.951Frequency(GHz)84868890929496Degree(°)Phase difference between S31 and S4190°0.80.850.90.951Frequency(GHz)84868890929496Degree(°)Phase difference between S32 and S4190°Figure 3.10 The illustration of peak search algorithm of Demeter for polarization alignment
calibration.

phase index that enables polarization alignment, namely obtaining the highest SNR at the gateway

side. Intuitively, a LoRa node can send 𝑛 probing beacons that cover all available phase indexes

to a LoRa gateway. Then, the gateway feedbacks the one with the highest SNR to the node. For

example, in Demeter prototype, we have 45 phase indexes. Figure 3.10 shows the SNR changes

(e.g., blue dots) for a cross-soil link deployed in the real world. We can see phase index 13 with the

highest SNR 4.6 dB should be selected.

Heuristic Calibration Algorithm: Demeter transmits beacons with varied polarizations in a short

duration. The cross-soil communication links are quite stable during a short period [150, 151].

Therefore, in Figure 3.10, the SNR changes comply with a sine wave approximately, where only

one valley and one peak exist. By noticing this property, we can reduce the transmission number

of probing beacons of current SF in a heuristic way. The algorithm contains two parts. First, we

use six uniformly distributed phase indexes to determine the optimal phase index’s range. Then, we

use binary search to find the optimal phase index in this range iteratively. For example, Figure 3.10

shows that the LoRa node first transmits six probing beacons using the phase indexes marked by

the red circles. The gateway feedbacks that the optimal phase index is in the range of either [1,2]

or [2,3]. Then, the node sends another two probing beacons with phase indexes 1’ (i.e., the middle

of [1,2]) and 3’ (i.e., the middle of [2,3]). The gateway feedbacks that the new search ranges [1’,2]

61

010203040Phase Index-10-505SNR(dB)1231'3'First IterationSecond Iterationand [2, 3’]. The node and gateway repeat this process to find the optimal phase index. In practice,

we control the number of iterations to balance the polarization accuracy and the energy cost at the

node side.

3.3.5 Link Calibration Scheduling

In our empirical study in Section 3.2.1, we observe that the SNR of a cross-soil LoRa link is

influenced by the changes in soil conditions (e.g., moisture, temperature) significantly. When soil

conditions change, the SNR is also degrading or fluctuating. With this observation, we need to

trigger link calibration more frequently if soil conditions change fast. In Demeter, LoRa nodes

and gateways periodically trigger the polarization alignment calibration Θ times in a day. If the

soil condition (e.g., rainfall, snowfall, fertilization, irrigation) changes frequently, we set a large Θ.

Otherwise, we keep Θ small to reduce the energy costs. Therefore, based on the weather report,

agriculture activity plan, and soil sensory data, we empirically configure Θ every day.

3.4

Implementation

We implement a prototype with three COTS hardware modules and a customized hybrid cou-

pler operating on the US915 band. The COTS hardware modules include a 3 dB XRDS-RF 698-

2700MHz passive RF splitter [132], an ANT627-NF-PANEL-MIMO-OD [133] dual-polarized an-

tenna, and a digital phase shifter PE44820 [131] with 45 available phase offset indexes. Considering

the loss of the phase shifter, we apply another identical one with no phase offset to the other signal,

balancing the strength of the two split signals.

We implement the hybrid coupler on two kinds of substrate: FR4 with 1.6 mm thickness and

Rogers5880 with 0.5 mm thickness. The 0402 patch capacitors with values of 4.3pF and 3.3pF

are used in Demeter on the Rogers5880 and FR4 coupler, respectively. As shown in Figure 3.11b,

the length and width are 63 mm and 25 mm, which is a relatively compact size for the IoT device

working in sub-1G frequency. We use a vector network analyzer to measure the S parameters of
our hybrid couplers. The phase difference between 𝑆31 and 𝑆41 is 89.6(cid:14) and the phase difference
between 𝑆32 and 𝑆42 is 90.7(cid:14). The signal loss of Demeter is 1.5 dB. The extra transmission power
consumption of Demeter is 600 𝜇W, which is neglectable compared with COTS LoRa nodes’ tens

62

(a) Implementation.

(b) Coupler

(c) Experiments deployment on a wheat farm

Figure 3.11 The implementation of Demeter prototype and experiment environments. The box
containing a battery, a LoRa node, and our prototype of Demeter RF components is buried in loam
soil at the top right corner of subfigure (c). Moisture experiments with an additional moisture sensor
are shown in the bottom right corner of (c).

of mW transmission power. Moreover, since we use a wide-band dual-polarized antenna and phase

shifter, the total cost of the Demeter prototype is 150 US dollars. We can lower the cost to about

20 US dollars by integrating all these modules into a unified circuit with a US915 band RF-front.

The cost can be further lowered with volume production.

As Figure 3.11a shown, we implement a Demeter node by connecting Demeter to a COTS LoRa

node with Semtech SX1276 radio chip [134]. For LoRa gateways, we use UHD and GNUradio to

control a USRP N210 [152] with an SBX 400-4400 RF daughterboard [153]. We also use a COTS

LoRa gateway with Semtech SX1302 radio chip [112] and SoilWatch 10 moisture sensor [140]. We

63

Gateway: USRP N210SX1302 LoRa GatewayMoisture Sensor:PINO-TECH Soil Watch 10COTS LoRa node:DraginoShield Arduino UNO Indoor ContainerRogers5880FR4AbovegroundGatewayUnderground Demeter in boxAbovegroundGatewayDemeter NodeGatewayMoisture SensorLoam Soilimplement the polarization alignment calibration algorithm using the standard LoRaWAN Class A

MAC, which is the default and most energy-efficient one. We run a heuristic algorithm script at the

gateway to evaluate the SNR from the LoRa node in soils. We set RECEIVE_DELAY as 10 s to

give enough response time for the gateway to generate feedback. We adopt 2 iterations and send 7

probing beacons in total in each calibration process. At the receiving window, the LoRa node turns

off the phase shift function but keeps receiving signals through the phase shifter, which does not

influence the signal reception.

3.5 Evaluation

We evaluate the communication reliability of Demeter in both indoor and in-field environments.

In addition, we demonstrate the energy efficiency of Demeter regarding different calibration trigger

frequencies.

Performance Metrics: We use two metrics to evaluate the performance of Demeter. The first is

SNR of the received packets, which is a widely used metric in wireless communication to indicate

communication reliability. The higher the observed SNR is, the more reliable the communication

is [154]. All the SNR gain values are net gain. The second is Energy Consumption per Day. We

use Joule per day (J/Day) to present energy consumption under various settings.

Baseline Method: We compare Demeter with the default linear polarized antenna without polar-

ization alignment. In our implementation, we only use one dipole antenna of the dual-polarized

antenna to generate the baseline signal with the same power as Demeter.

3.5.1

Indoor Experiment

Experiment Settings: We conduct indoor experiments to verify the SNR gain of Demeter in polar-

ization alignment for cross-soil communication. We bury a Demeter node in a plastic container of

sandy soil and utilize the COTS LoRa gateway to receive packets as Figure 3.11a shown. The dis-

tance between the antenna and the gateway is 20 m and the in-soil distance between the antenna and

the edge of the container is 0.6 m. Then we gradually add water and stir the soil evenly to change

VWC from 0% to 48%. For Demeter, we calibrate the polarization, and the antenna orientation is

randomly placed, then we calculate the average SNR of 10 packets. For the baseline method, we

64

Table 3.2 Comparison between COTS LoRa devices and Demeter performance in sandy soil with
different moisture levels. The metric is SNR with unit dB.

VWC (cm3/cm3)
Max SNR
Min SNR
Demeter SNR

0% 8% 16% 24% 32% 40% 48%
-6.1
5.9
-12.7
-5.9
-5.9
6.1

-2.6
-10.6
-2.1

5.2
-6.5
5.1

5.7
-7.2
6.0

3.2
-7.7
3.3

0.2
-9.6
0.5

collect and estimate the SNR of 10 packets from the LoRa node for each of the 16 different ori-

entations of the manually rotated antenna of the LoRa gateway. For each VWC level, we keep the

same wireless link between the two methods by switching the transmitter connected to the antenna

ports without changing soil conditions. The spreading factor (SF) of LoRa transmission is 10.

Results: Table 3.2 shows the maximum (i.e., Max SNR) and minimum (i.e., Min SNR) SNR values

of the baseline method at different antenna orientations and the average SNR value of Demeter

at different VWC levels. We can see that the Demeter’s SNR is 0.21 dB higher than Max SNR

on average. Moreover, Demeter SNR can achieve 11.6 dB SNR improvement compared to Min

SNR. The results indicate that Demeter can achieve good polarization alignment to maintain high

communication quality compared to manual polarization alignment adjustment.

3.5.2

In-field Experiment

We evaluate the performance of Demeter under five different configurations in terms of environ-

ment and deployment outdoors. We conduct experiments on a wheat farm as shown in Figure 3.11c.

In the experiments, we dig a hole and bury a battery, a LoRa node, and our prototype of Demeter

RF components (as mentioned in Section 3.3) in a plastic box to protect devices. Considering the

potential ultra-low SNR outdoors, we utilize the USRP as the gateway to collect raw signals and

compute average SNR among multiple symbols. After the polarization alignment calibration, we

collect 12 packets to measure the average SNR value for Demeter. For the baseline, we also collect

12 packets to measure the average SNR. We set a 45(cid:14) inclined antenna (Figure 3.14) of the gateway

towards to underground node in baseline and Demeter. The default moisture is 11%, and the default

soil type is loam soil.

1) Horizontal Distance Experiment Settings: As shown in Figure 3.11c, we conduct experiments

65

(a) SF=7

(b) SF=8

(c) SF=9

(d) SF=10

Figure 3.12 Impact of horizontal distance to Demeter.

in multiple horizontal distances from 50 m to 350 m. Orange triangles represent the locations of

the gateway at various distances. The blue star is the location of the buried node. The buried depth

is 35 cm. We set different SFs from 7 to 10 in the experiments.

Results: As Figure 3.12b and Figure 3.12a shown, compared to the baseline, the SNR gain of

Demeter is at least 8.87 dB and 9.94 dB with SF-8 at 150 m and SF-7 at 100 m respectively. The

average SNR gain of SF9 and SF10 are 6.16dB and 6.55dB, respectively. The average SNR gain of

Demeter achieves 6.86 dB across different configurations. We can observe that the SNR difference

between Demeter and the baseline varies because the polarization of the baseline is random when

66

050100150200250300350Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold050100150200250300350Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold050100150200250300350Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold050100150200250300350Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold(a) SF=7, Distance=100m

(b) SF=8, Distance=100m

(c) SF=9, Distance=100m

(d) SF=10, Distance=300m

Figure 3.13 Impact of depth to Demeter.

Demeter achieves polarization alignment. Due to terrible polarization, the baseline signals may

drop lower than the required SNR for successful decoding. Standard LoRa may increase SF to

adapt. Instead, Demeter can recover wireless communication without varying SF. In Figure 3.12,

if the distance is beyond 250 m, the baseline suffers from misalignment polarization, and Demeter

trigger polarization alignment calibration to improve the signal SNR from under threshold to above

threshold. Every time SF is added by one, the distance of Demeter can be enlarged by 50 m. For

SF7, SF8 and SF9, distance promotion is 4(cid:2), 2.5(cid:2), 2(cid:2) and 1.75(cid:2) respectively in the Figure 3.12a,

Figure 3.12b , Figure 3.12c and Figure 3.12d. Based on the results, Demeter can achieve 2.5(cid:2)

67

152535Depth(cm)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold152535Depth(cm)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold152535Depth(cm)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold152535Depth(cm)-20-15-10-50510SNR(dB)DemeterBaselineSNR thresholdFigure 3.14 The three different antenna orientations.

communication range on average compared with COTS LoRa. It is equivalent to a 6.56(cid:2) coverage

area gain.

2) Depth Experiment Settings: To verify the performance of Demeter in different underground

depths, we placed the Demeter node into three depth levels underground, 15 cm, 25 cm, and 35 cm,

as SF is set from 7 to 10. We evaluate Demeter under the horizontal distance as 100 m and 300 m.

Results: Figure 3.13a, Figure 3.13b, and Figure 3.13c show the results. In all settings, we can see

the SNR gain of Demeter compared to the baseline. Under 300 m horizontal distance and SF-10,

the baseline signal can be detected only when the depth is 15 cm. Demeter make the LoRa signal

appear again when depth is 25 cm or 35 cm as shown in Figure 3.13d. The depth of Demeter can

be extended to 2.33(cid:2) compared with the baseline. In Figure 3.13b, the SNR gain is only 2.2 dB

at 25 cm depth while the SNR gain is 8.67 dB with 35 cm depth. This indicates the wireless

link is unstable when depth changes. The average SNR gain of Demeter achieves 5.30 dB across

different configurations. Besides horizontal distance, Demeter can also improve maximal soil depth

by polarization alignment calibration for cross-soil wireless communication.

3) Antenna Orientations Experiment Settings: The different antenna orientations of the gateway

can influence the polarization alignment directly. In this experiment, the gateway antennas with

vertical and horizontal orientations are selected to explore the impact of antenna orientation for

Demeter. SF is set as 10, and the depth is 35 cm. Figure 3.14 shows the three kinds of antenna

orientation.

Results: The measurement results are illustrated in Figure 3.15. We can observe that compared to

the baseline, the average SNR gain of Demeter in the vertical antenna (Figure 3.15a) is 6.41 dB,

68

Vertical45°InclinedHorizontal(a) vertical

(b) horizontal

Figure 3.15 Impact of antenna orientations to Demeter.

which is close to the 6.34 dB SNR gain of the horizontal antenna (Figure 3.15b), and 6.55 dB of 45(cid:14)

inclined antenna in Figure 3.12d. This verifies that even if the gateway antenna is put in different

orientations suffering from beam mismatching and polarization plane changes, Demeter can work

perfectly to provide reliable communication.

4) Moisture Experiment Settings: We explore the impact of different moisture levels (11%, 16%,

25%, 36%, 47%). We use a moisture sensor to measure the VWC before we bury Demeter node

underground. The measurement depth of VWC sensor and Demeter node is 35 cm. SF is set as 10.

Results: As illustrated in Figure 3.16, the average SNR gains with 16% VWC and 25% VWC

are 6.41 dB and 6.24 dB respectively compared to the baseline. Both values are close to 6.55 dB

gain in Figure 3.12d with VWC 11%. The SNR gains decrease slightly due to VWC growing up

as the attenuation also increases. The SNR gains in Figure 3.16b range from 3.54 dB to 8.64 dB

while SNR gains range from 4.25 dB to 8.49 dB in Figure 3.16a. With higher moisture levels in

Figure 3.16c and Figure 3.16d, the overall SNR decreases because of the moisture-caused attenua-

tion. But Demeter can still compensate 5.29 dB SNR on average when VWC=36% and enlarge 4(cid:2)

horizontal distance in flooding status with VWC=47%. This indicates Demeter can accommodate

different soil environments with various moisture conditions to achieve polarization alignment.

5) Soil Type Experiment Settings: To explore the influence of soil type, we evaluate our system

69

050100150200250300Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold050100150200250300Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold(a) VWC=16%

(b) VWC=25%

(c) VWC=36%

(d) VWC=47%

Figure 3.16 Impact of moisture to Demeter.

under three agricultural soil types: loam soil, clay soil, and sandy loam soil. The percentages of

silt, sand, and clay can categorize soil. Loam has balanced components and good fertility. It holds

moisture and nutrients well, drains excess water easily, and allows air to circulate around plant roots,

which is ideal for gardening and agriculture. Clay takes the most percentage of clay soil, which has

smaller and fewer pore spaces than loam soil, which means that water drains slowly and can cause

waterlogging and compaction. Sandy loam soil has more sand compared to loam. It has a coarse

texture and a loose structure that allows water and air to flow through easily. In the experiments,

the depth is 35 cm and SF is set as 8.

70

050100150200250300350Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold050100150200250300350Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold050100150200250300350Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold050100150200250300350Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold(a) Sandy Loam Soil

(b) Clay Soil

Figure 3.17 Impact of soil type to Demeter.

Results: The results for loam soil are shown in Figure 3.12b. Figure 3.17 shows the results for

sandy loam and clay soils. Demeter achieves higher SNR in sandy loam of Figure 3.17a than loam

with the same horizontal distance while Demeter achieves lower in clay soil of Figure 3.17b than

in loam. Baseline signals in sandy loam yield the best performance, while clay soil ones are the

worst. The SNR difference between these two types of soil is 3.39 dB. The SNR gain is 6.9 dB,

6.93 dB, and 7.08 dB in loam, sandy loam, and clay soils, respectively. This indicates Demeter can

work in multiple types of soils and achieve good communication quality and similar SNR gains.

3.5.3 Energy Consumption Analysis

Experiment Settings: By leveraging the SNR gain of Demeter, we can increase the data rate

with a lower SF to save energy compared to the baseline under the same link budget. Previous

experiment results indicate Demeter can achieve at least 5 dB SNR gain, which means we can reduce

SF by 2 grades with a lower SNR threshold compared to the standard LoRa [40,100,134,155]. For

example, if the standard LoRa has to use SF12, we can use SF10 for packet transmission. Given

different link budgets and SF settings, the energy saving for each packet ranges from 0.0569 J to

0.5786 J with 3.3 V voltage.

First, we make a LoRa node sending six packets per hour to evaluate the energy efficiency

of Demeter compared to the standard LoRa under different polarization calibration frequencies.

71

0255075100125150Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR threshold0255075100125150Distance(m)-20-15-10-50510SNR(dB)DemeterBaselineSNR thresholdFigure 3.18 The analysis of energy consumption per day under various settings.

Based on the fact that significant fluctuations in SNR occur only 5 times over a 9-day period in

Figure 3.3, we set Θ from 0.5 to 6. Specifically, Demeter 1 triggers the calibration once every

two days. Demeter 2, 3, and 4 trigger the calibration 2, 4, and 6 times per day. By default, the

payload size is set to 26 bytes. We calculate the energy consumption per day under four different

link budgets from SF9 to SF12.

To further verify the performance of various payload lengths, we adopt six payload sizes from

10 bytes to 30 bytes. For each payload size, we calculate the average energy consumption per day

of 144 packet transmissions under link budgets from SF9 to SF12 (i.e., 36 packets for each link

budget). We compare Demeter 2 with the standard LoRa and Fixed-SF12, which always applies

SF12 to transmit packets no matter the current link budget.

Results: In Figure 3.18a, Demeter 1–4 outperforms the standard LoRa under different link budgets.

For example, compared to the standard LoRa, Demeter 2 reduces energy consumption per day by

68.3% and 57.9% for SF12 and SF9, separately. Moreover, we can see that energy consumption per

day increases when calibration trigger frequencies increase. Demeter 1’s average energy consump-

tion per day under the four SF settings is 34.1% of the standard LoRa while Demeter 4’s increases

to 48.5%. Figure 3.18b illustrates the energy consumption under different payload sizes. Demeter

2 reduces energy consumption per day by 82.0% compared to Fixed-SF12 and 63.4% compared

to the standard LoRa. For a LoRa node with a 4,400 mAh energy store (i.e., two mid-grade Al-

kaline AA batteries), this would extend the system lifetime roughly 5.5(cid:2). The results indicate the

72

SF9SF10SF11SF12Spreding Factor020406080100120Energy (J/Day)Standard LoRaDemeter1Demeter2Demeter3Demeter4101418222630Payload Size (bytes)0306090120150Energy (J/Day)Standard LoRaDemeter2Fixed-SF12energy benefits of polarization alignment are far more than the energy cost incurred by alignment

calibration in Demeter.

3.6 Related Work

Cross-soil IoT Systems: Ad-hoc Wireless Underground Sensor Networks (WUSN) [85,91,93,156,

157] is well studied for cross-soil IoT. For example, Wang et al. [156] propose a soil measurement

system based on WUSN. Dong et al. [158] and Silva et al. [85] present autonomous precision irriga-

tion systems with WUSN. Compared to ad-hoc WUSN, we focus on cross-soil LPWAN, a parallel

IoT paradigm. Moreover, some works [88, 89, 159] utilize COTS LoRa and NBIoT to demonstrate

the feasibility of cross-soil LPWAN. In comparison, Demeter further enhances the reliability, com-

munication range and depth, and energy efficiency of LoRa cross-soil communication.

In-soil Communication Method: Some works [90, 91, 160, 161] focus on extending the depth a

sensor can be buried at with electromagnetic waves [91, 161], magnetic coils [90], and acoustic

signals [160]. For example, Sun et al. [90] adopt magnetic induction to reduce signal path loss

in soils. The communication system [160] can achieve up to 50 m buried depth with a low-cost

acoustic transceiver. Although magnetic and acoustic signals are able to tolerate the signal loss

brought by soils, they cannot enable long-range communication over the air, which hampers the

farm-scale agricultural data collection needs. In contrast, Demeter is able to achieve much longer

cross-soil communication distance for general-purpose agricultural IoT systems.

Polarization-aware Communication for IoT: LLAMA [115] implemented a reconfigurable meta-

surface to rotate the polarization of signal between transmitter and receiver. Polartracker [110]

and Polarscheduler [111] propose a polarization-aware link model to communicate in polarization-

aligned periods. In comparison, Demeter enables adaptive polarization alignment by a customized

RF front-end for cross-soil LPWAN with low protocol overhead.

3.7 Discussion and Future Work

VWC range in our field study. Our in-field experiments were conducted just after a wheat

harvest. Therefore, no irrigation was applied, leading to lower VWC values than those reported

in the growing season [162–165] in our field study. Although the controlled experiments (e.g.,

73

Table 3.2, Figure 3.16) demonstrate a consistent performance gain when the VWC increases to

48%, we plan to conduct a large-scale deployment in the next growing season to evaluate Demeter

performance in a more realistic VWC.

Hobbyist soil moisture sensor. The soil moisture sensor [140] we used to measure the groundtruth

VWC is a hobbyist-level sensor. Although it can correctly report the trend of soil moisture changes,

the accuracy of soil moisture measurement is not guaranteed. Since the specific characteristics of

the soil greatly influence the accuracy of soil moisture sensors, a more accurate sensor with a soil-

specific calibration would be used. In our future work, we will employ a calibratable and accurate

professional-level soil moisture sensor for exploring soil-specific polarization

Adaptive polarization calibration for large-scale deployment One of our future works is to go

beyond the physical layer design in this paper to build a reliable and energy-efficient network stack

for achieving large-scale deployment. Considering environment and soil diversity in large-scale

deployment, different LoRa nodes may need to be calibrated with different schedules. Instead of

our periodical polarization calibration, on-demand polarization calibration will play a significant

role in large-scale deployment as it allows a LoRa node to calibrate its polarization adaptively. A

machine learning based method may be feasible to determine the trigger for a polarization calibra-

tion by taking multi-dimensional environmental and soil property changes as inputs. However, it

is challenging to balance inference accuracy and computation cost on energy-limited LoRa nodes.

Moreover, we will integrate RF front-end circuit design of Demeter to facilitate large-scale deploy-

ment and quantify the performance and cost benefits compared to other cross-soil communication

solutions.

3.8 Conclusion

In this work, we present Demeter, a novel system that enables reliable cross-soil LPWAN com-

munication. The main contribution of Demeter is a low-cost antenna system that aligns the po-

larization between LoRa node and gateway antennas and works with COST single RF-chain LP-

WAN radios. We also develop a low-cost hybrid coupler to encode and change polarization degrees

with phase information. Furthermore, we propose a heuristic algorithm to calibrate the polarization

74

alignment automatically. We implement Demeter with customized PCB circuits and COTS devices

to evaluate its performance in various soil conditions and scenarios. Demeter can work with vari-

ous soil types and different environmental conditions. The results show that Demeter can achieve

up to 9.94 dB SNR gain outdoors and 11.6 dB indoors, 4(cid:2) horizontal communication distance, at

least 20cm underground depth improvement, and up to 82% energy reduction compared with COTS

LoRa node.

75

CHAPTER 4

PRISM: HIGH-THROUGHPUT LORA BACKSCATTER WITH NON-LINEAR CHIRPS

4.1

Introduction

LoRa [1, 7] is known for its low-power and long-range communication enabling low-power

wide-area network (LPWAN) for massive IoT, which has been seen as an important part of the next-

generation networks [8]. LoRa uses Chirp Spread Spectrum (CSS) modulation, in which a LoRa

symbol is represented by a linear chirp. During the demodulation (called “dechirp”), the energy

of the whole chirp will be accumulated at a signal frequency in the spectrum to enable reliable

decoding even when the single strength is below the noise floor. By leveraging the advantage of

LoRa’s noise tolerance, some works [166–169] show the designs which adopt LoRa’s chirps in

backscatter systems to achieve long communication distance (e.g., 1.1 km to 2.8 km) benefiting the

deployment of the backscatter systems in practice [170]. On the other hand, backscatter tags usually

transmit data without a time backoff. Concurrent LoRa backscatter transmissions in a large-scale

system increase the collision probability. As a result, the system reliability might be significantly

degraded [169]. How to resolve the signal collision is a critical issue in applying LoRa backscatter

in general.

Existing works [168,169] addressed this problem by creating more orthogonal logical channels

in the frequency domain, allowing multiple backscatter tags to transmit concurrently by selecting

different logical channels. For example, NetScatter [169] can decode parallel packets with dis-

tributed chirp spread spectrum coding and a specially designed excitation signal. P2LoRa [168]

achieves parallel decoding by shifting the ambient LoRa chirps with different frequencies. The

evaluation results show that 256 and 101 concurrent transmissions can be realized by NetScatter

and P2LoRa, respectively. However, these methods cannot step further to support over 500 or more

concurrent transmissions efficiently due to the LoRa’s narrow-band nature.

Recently, CurvingLoRa [171] has shown the possibility of exploiting non-linear chirps to enable

high-throughput LoRa concurrent transmission. First, different types of non-linear chirps can be

used to transmit data concurrently [172]. Second, the transmissions with the same type of non-

76

Figure 4.1 The illustration of Prism. The tags take linear chirps as excitation sources and backscatter
different types of non-linear chirps for concurrent transmission.

linear chirp can be concurrently decoded when the timing of these transmissions is not exactly

aligned [171]. This motivates us to rethink the design of the LoRa backscatter system with non-

linear chirps.

In this paper, we propose Prism, a high-throughput LoRa backscatter method by embracing the

non-linear chirps to enable concurrent transmissions. As shown in Figure 4.1, a commercial-off-

the-shelf (COTS) LoRa node generates linear chirps (e.g., the dashed brown chirp) as excitation

signals. When multiple (e.g., three) Prism tags detect the excitation signals, they transform the lin-

ear chirps to different non-linear chirps (e.g., the solid blue, green, and red chirps) to modulate their

information. When a gateway receives these signals with the different chirp types, we can easily

distinguish the excitation and backscatter chirps to decode the information of different tags [171].

However, converting a linear chirp to its non-linear counterpart with limited resources on a

backscatter tag is not trivial. To enable an efficient chirp-type transform, we must address the

following challenges:

First, frequency shifts between an excitation linear chirp and a backscatter non-linear chirp

change over time. It is challenging to generate the time-variant frequency shifts to transform the

77

Non-linear Chirp GenerationGatewayLoRa NodeLoRa nodePrism tagQuadraticQuadraticCubicLinearGatewayPrism tagchirp types in a lightweight manner. In Prism, we use a timer to trigger different frequency shifts

at different times. When the timer interrupt is triggered, a Micro Controller Unit (MCU) generates

an integer input indicating the targeted frequency shift at the current time to a Digital Analog Con-

verter (DAC). Then the DAC outputs a specific voltage, which is the input of a Voltage Controlled

Oscillator (VCO). At last, the VCO outputs a square wave signal with the targeted frequency, which

controls a radio frequency switch to change the antenna impedance/density to add the specific fre-

quency shift. In this way, we keep the energy consumption of the chirp-type transform low.

Second, it is not trivial to calculate the frequency shifts accurately. In Prism, we use a frequency-

shift function to calculate the frequency shifts over time. Specifically, all chirps have their math

functions, which can be linear, quadratic, cubic, or higher-order, to shape different chirp types.

We can accurately calculate the frequency shifts between two different types of chirps by using

their math function in the time-frequency domain. Hence, we use the same way to generate the

frequency-shift function between the linear function and the corresponding non-linear function.

Third, due to various noises, how to reliably demodulate backscatter non-linear chirp symbols

is the last challenge. In Prism, we observe that for a backscatter non-linear chirp symbol, the whole

chirp energy will be accumulated at two peaks on the spectrum after a non-linear dechirp process.

We develop a phase search method to combine the two peaks coherently to enhance the received

signal strength. Moreover, a misaligned non-linear chirp suffers from severe frequency leakage. We

utilize pilot linear chirps in an excitation LoRa packet to generate pilot non-linear chirps to calibrate

such frequency offsets. In these ways, we can demodulate backscatter chirp symbols reliably.

We implement Prism with a COTS LoRa device, a customized PCB circuit, and a USRP. We

evaluate its performance in outdoor and indoor scenarios. The results show that Prism achieves

the highest throughput of 68.36 Kbps when 700 tags concurrently transmit, indicating a 6(cid:2) higher

transmission concurrency than state-of-the-art. In addition, in the same channel, the bit error rate

of the seven concurrent transmissions is less than 1% by using seven different types of non-linear

chirps. The contributions of this paper are summarized as follows:

• We first introduce non-linear chirps to LoRa backscatter systems. The system scalability is

78

Figure 4.2 The illustration of LoRa CSS modulation and demodulation with the initial frequency
shifted up-chirp and the base down-chirp.

significantly improved with our enhanced network throughput.

• On resource-limited backscatter tags, we propose a lightweight and low-power method to

convert a linear excitation chirp to non-linear ones, which can be reliably decoded at a gate-

way.

• We prototype Prism and evaluate its performance in real environments. The results show the

highest throughput of 68.36 Kbps with 700 tags. It is 6(cid:2) higher than state-of-the-art in terms

of transmission concurrency.

4.2 LoRa Preliminary

In this section, we briefly introduce LoRa CSS modulation and explain the background knowl-

edge of the LoRa concurrent transmission enabled by non-linear chirps.

4.2.1 LoRa CSS Modulation

A kind of CSS mechanism is adopted by LoRa. Given a bandwidth 𝐵𝑊, as shown in Figure 4.2,

the basic communication unit of LoRa CSS modulation is a linear base up-chirp (e.g., the solid blue
line in the left sub-figure) whose frequency increases linearly from (cid:0) 𝐵𝑊
be donated as 𝐶 „𝑡” = 𝑒 𝑗2𝜋„(cid:0) 𝐵𝑊

2 over time, which can
”𝑡. The key character of the CSS modulation is that a time delay

2 to 𝐵𝑊

‚ 𝑘𝑡
2

2

in a chirp can be transformed into a frequency shift. Therefore, encoded data bits can be modulated

by the initial frequency offset or the cyclic time shift in the base up-chirp (e.g., the solid blue line in

79

+𝐁𝐖𝟐−𝐁𝐖𝟐𝐟𝑡02𝑆𝐹−1𝐅𝐅𝐓𝐛𝐢𝐧+𝐁𝐖𝟐−𝐁𝐖𝟐𝐟𝑡02𝑆𝐹−1𝐅𝐅𝐓𝐛𝐢𝐧FFT.AbsFFT.AbsFigure 4.3 The illustration of the decoding failure due to the collision between the linear chirp
symbols from different backscatter tags.

the right sub-figure). A chirp symbol can carry multiple bits of information, which is determined

by the spreading factor (SF). By combining symbols with CSS modulation, a LoRa packet consists

of many chirp symbols, encoding multiple pieces of information.

The process of demodulation is called “dechirp”. During the dechirp, it will multiply a received

chirp symbol with a base down-chirp, which is the conjugate of the base up-chirp (e.g., the dashed

orange line in Figure 4.2). After FFT (Fast Fourier Transform), a peak will appear at an FFT

frequency bin, showing the initial frequency of the received chirp symbol. Therefore, if a LoRa

configuration defines 𝑁 different initial frequency offsets, its SF equals to 𝑙𝑜𝑔2𝑁.

4.2.2 Concurrent Transmission in LoRa Backscatter

Collision happens when multiple concurrent LoRa backscatter packets arrive at a gateway in a

close time. The time difference comes from the different propagation delays of the multiple tags.

Usually, after detecting the identical excitation LoRa signals, the arrival time difference of the

concurrent backscatter chirp symbols at a gateway is less than a microsecond [168]. As shown in

Figure 4.3, assume three tags use linear chirps to modulate data in the same channel. When we apply

the dechirp on the overlapped linear chirp symbols, we will observe three energy peaks at close FFT

bins on the spectrum. For a tag (e.g., blue linear chirp), we cannot reliably detect its energy peak if it

is overwhelmed by the interference energy peaks on the spectrum. This situation will happen when

the linear chirp symbol may experience a larger attenuation than the interference chirp symbols due

to the near-far issue [171]. As a result, we inevitably have decoding ambiguity with two overlapped

80

+𝑩𝑾𝟐−𝑩𝑾𝟐𝒇𝒕𝑭𝑭𝑻.𝒂𝒃𝒔𝟎𝟐𝑺𝑭−𝟏𝑭𝑭𝑻𝒃𝒊𝒏Cannot distinguishFigure 4.4 The illustration of the collision resolving between different types of linear and non-linear
chirps from different backscatter tags.

packets in concurrent LoRa transmissions. Only the stronger packets are possibly decoded, but they

will suffer more from the decoding ambiguity with massive concurrency.

To resolve the LoRa collision problem, CurvingLoRa [171] defines a non-linear base up-chirp

whose frequency increases non-linearly from (cid:0) 𝐵𝑊
2

to 𝐵𝑊

2 over time. There are multiple types of

non-linear chirps, like quadratic, quartic, or trigonometric. As shown in Figure 4.4, we have three

backscatter tags that use linear (red), quadratic (blue), and quartic (brown) chirps to modulate data,

respectively. Their backscatter chirps overlap with each other. For demodulation, if the initial

frequency offset of the interference chirp symbol is 𝑓0, applying the dechirp on the corresponding

chirp symbol can be described as follows:

𝑒 𝑗2𝜋„ 𝑓0‚ 𝑓 „𝑡””𝑡 (cid:2) 𝑒(cid:0) 𝑗2𝜋 𝑓 „𝑡”𝑡 = 𝑒 𝑗2𝜋𝐹 „𝑡”𝑡

(4.1)

where 𝑓 „𝑡” is a time-frequency function to depict the shape of a base up-chirp, and (cid:0) 𝑓 „𝑡” indicates

the conjugate base down-chirp. 𝐹 „𝑡” is a dechirp frequency function to depict the resulting signals.

In Figure 4.4, the 𝑓0 is 0, and the time-frequency functions of linear, quadratic, and quartic chirps

can be expressed as 𝑓𝑙𝑖𝑛𝑒𝑎𝑟 „𝑡” = 𝑎1𝑡, 𝑓𝑞𝑢𝑎𝑑𝑟𝑎𝑡𝑖𝑐 „𝑡” = 𝑏1𝑡2, 𝑓𝑞𝑢𝑎𝑟𝑡𝑖𝑐 „𝑡” = 𝑐1𝑡4, respectively.

Taking the demodulation of the quadratic non-linear chirps as an example. The quadratic base

down-chirp is the green curve in Figure 4.4. When we apply the quadratic dechirp on the quadratic

chirp, we obtain 𝐹𝑞𝑢𝑎𝑑𝑟𝑎𝑡𝑖𝑐 „𝑡” as follows:

𝐹𝑞𝑢𝑎𝑑𝑟𝑎𝑡𝑖𝑐 „𝑡” = 0

(4.2)

81

+𝑩𝑾𝟐−𝑩𝑾𝟐𝒇𝒕𝑭𝑭𝑻.𝒂𝒃𝒔𝟎𝟐𝑺𝑭−𝟏𝑭𝑭𝑻𝒃𝒊𝒏𝑝𝑟𝑝𝑖Correct Demodulation:𝑝𝑟>𝑝𝑖which will be an energy peak at FFT bin 0, as we expected. However, when we apply the quadratic

dechirp on the linear and quartic chirps, we obtain 𝐹𝑙𝑖𝑛𝑒𝑎𝑟 „𝑡” and 𝐹𝑞𝑢𝑎𝑟𝑡𝑖𝑐 „𝑡” as follows:

𝐹𝑙𝑖𝑛𝑒𝑎𝑟 „𝑡” = 𝑎1𝑡2 (cid:0) 𝑏1𝑡3

𝐹𝑞𝑢𝑎𝑟𝑡𝑖𝑐 „𝑡” = 𝑐1𝑡5 (cid:0) 𝑏1𝑡3

(4.3)

(4.4)

We can see that instead of the constant frequency in 𝐹𝑞𝑢𝑎𝑑𝑟𝑎𝑡𝑖𝑐 „𝑡”, the frequencies of 𝐹𝑙𝑖𝑛𝑒𝑎𝑟 „𝑡”

and 𝐹𝑞𝑢𝑎𝑟𝑡𝑖𝑐 „𝑡” are time-variant with a non-zero 𝑡, leading to the frequency changes over time.

Therefore, the energy of the linear and non-linear interference chirp symbols is no longer accumu-

lated to high energy peaks on the spectrum. As shown in Figure 4.4, the energy of linear and quartic

interference chirps (e.g., the small red and brown dashed energy noises) spreads over multiple, clus-

tered FFT bins on the spectrum. This scattering effect makes the blue symbol can be demodulated

successfully with an accumulated energy peak (e.g., the blue solid energy peak in Figure 4.4) on the

spectrum. Then, we can utilize specific types of dechirp windows to demodulate LoRa backscatter

packets with specific chirp types in concurrent transmissions of the multiple tags. Thus, the energy

scattering effect among different types of non-linear chirps enables multiple quasi-orthogonal log-

ical channels. The collision tolerance ability of non-linear chirps motivates us to develop a LoRa

backscatter involving non-linear chirps with higher network throughput.

4.3 System Design

4.3.1 System Overview

The system overview is shown in Figure 4.5. We adopt the same tag hardware framework with

P2LoRa [168]. A Prism tag includes a preamble detection circuit, MCU, DAC, and VCO. The ul-

timate goal of our design is to enable 𝑁 resource-limited backscatter tags in the same channel to

capture an excitation linear chirp and transform it into 𝑁 orthogonal non-linear chirps, respectively,

which can be concurrently decoded at a gateway. To achieve the goal, the end-to-end system con-

sists of three parts: backscatter modulation, non-linear signal demodulation, and non-linear chirp

decoding. First, for a Prism tag, the preamble detection circuit wakes up its MCU once a linear

LoRa excitation packet is detected. Then the MCU will control the DAC and VCO to shift the

82

Figure 4.5 The overview of Prism system.

frequency of the excitation linear chirps to generate its non-linear chirps with a frequency-shift

function. Second, at the gateway, we calibrate various frequency offsets in the received non-linear

chirps and concurrently demodulate them with the dechirp by utilizing the energy scattering effect.

Finally, after demodulating the chirp symbols from different tags with different types of non-linear

chirps, we decode the bits embedded in these backscatter non-linear chirps from each tag.

4.3.2 Preamble Detection

Prism adopts a low-power preamble detection circuit [168, 173] to identify the arrival of linear

LoRa excitation packets. The circuit includes three parts: impedance matching, envelop shaping,

and preamble detection. The first two parts can capture the repetitive pattern of the continuous

base up-chirps in the preamble of a LoRa packet. The last part compares the captured pattern with

a template to determine whether linear LoRa excitation chirps are coming.

4.3.3 Non-linear Backscatter Chirp Generation

The first task of generating a non-linear backscatter chirp is to add frequency shifts on a linear

excitation chirp. As shown in Figure 4.6, after the linear excitation chirps are detected, we use the

MCU to control the output voltage of the DAC, which determines the frequency of square wave

signals generated by the VCO. The square wave signals are added on linear excitation chirps to

generate the non-linear backscatter chirps.

Specifically, the MCU uses 16-bit serial data to control the DAC. The upper 4 bits are control

83

BackscatterPacketsNon-linear GenerationPreambleDetectionBackscatterModulationRemoveOffsetConcentrateEnergyDemodulateRecoverBackscatterbitsDecodeAmbient LoRa signalBackscatter bitsN Backscatter PacketsExcitation SourceN Prism TagsGatewayFigure 4.6 The design of a lightweight frequency shift generation on a Prism tag.

bits, and the lower 12 bits give the input data. The output voltage 𝑉𝑂𝑈𝑇 of the DAC is given with

the following equation:

𝑉𝑅𝐸 𝐹 (cid:2) 𝑁
4096
where 𝑁 is the numeric value of the DAC’s binary input code, 𝑉𝑅𝐸 𝐹 is a constant value as the

𝑉𝑂𝑈𝑇 =

(cid:2) 𝑔𝑎𝑖𝑛

(4.5)

reference voltage, and the value of 𝑔𝑎𝑖𝑛 is 2, decided by the peripheral circuit. The output voltage

𝑉𝑂𝑈𝑇 of the DAC is the input of the VCO, generating a square wave signal, whose frequency 𝑓𝑂𝑈𝑇

can be calculated with the following equation:

𝑓𝑂𝑈𝑇 =

1𝑀 𝐻𝑧 (cid:2) 50𝑘
𝑁𝐷𝐼𝑉 (cid:2) 𝑅𝑉𝐶𝑂

(cid:2) „1 ‚

𝑅𝑉𝐶𝑂
𝑅𝑆𝐸𝑇

(cid:0)

𝑉𝑂𝑈𝑇
𝑉𝑆𝐸𝑇

”

(4.6)

where 𝑁𝐷𝐼𝑉 is 2, 𝑉𝑆𝐸𝑇 is the voltage on the SET pin, nominally set as 1.0V, and the 𝑅𝑉𝐶𝑂 and 𝑅𝑆𝐸𝑇

are two 100k resistors. With Equation 4.5 and 4.6, we can set a 16-bit integer value ranging from 0

– 65535 to 𝑁 to generate a square wave signal with the specific frequency 𝑓𝑂𝑈𝑇 . The square wave

signal further controls a radio frequency switch to add 𝑓𝑂𝑈𝑇 frequency shift on the linear excitation

chirps.

When a tag can shift the frequency over linear excitation chirps, the next step is to determine

the frequency shift needed at a specific time. In Figure 4.7, the left sub-figure illustrates the fre-

quency difference between a red linear base up-chirp and an orange non-linear base up-chirp (e.g.,

a quadratic shape), which are time aligned. We subtract the frequency of the linear base up-chirp

from that of the non-linear base up-chirp, defining the phase shifts needed over time, as shown in

the right sub-figure. For different linear excitation chirps, we use the same sequence of frequency

shifts to generate their non-linear counterparts for simplicity. In addition, the MCU uses a timer to

trigger different frequency shifts over time.

84

Preamble DetectorMicroControllerUnitDigital Analog converterVoltage Controlled OscillatorFigure 4.7 The time-frequency transform from a linear excitation to a non-linear chirp.

Intuitively, we can use a look-up table to store the frequency shifts at different time points.

However, in order to achieve an accurate non-linear chirp generation, we must store fine-grained

frequency shifts, which are hard to be supported with the limited storage space on the low-power

MCU. Instead of the look-up table, we compute the frequency shifts with an accurate time-frequency

function. We adopt the way proposed by CurvingLoRa [171] to define a non-linear base up-chirp
2 ) to ( 2𝑆𝐹
as a monotonic curve growing from (0,(cid:0) 𝐵𝑊
2 ). The coordinates 𝑥 and 𝑦 indicate time
and frequency, respectively. The monotonic curve function 𝑓𝑐 „𝑡” of a linear or non-linear chirp

𝐵𝑊 ,(cid:0) 𝐵𝑊

can be represented by the sum of a series of polynomial functions in the time-frequency domain as

follows:

𝑓𝑐 „𝑡” =

𝑛∑

𝑖=0

𝑘𝑖𝑡𝑖, 𝑡 2 »0, 2𝑆𝐹
𝐵𝑊

…, 𝑓𝑐 „𝑡” 2 »(cid:0)

𝐵𝑊

2

,

𝐵𝑊

2

…

(4.7)

Hence, for a linear base up-chirp, 𝑘0 = (cid:0) 𝐵𝑊

2 and 𝑘1 = 𝐵𝑊 2

2𝑆𝐹 . Its curve function is the following:

𝑓𝑙𝑖𝑛𝑒𝑎𝑟 „𝑡” = (cid:0)

𝐵𝑊

2

‚

𝐵𝑊 2
2𝑆𝐹

𝑡

(4.8)

In addition, to depict a quadratic non-linear base up-chirp, 𝑘0 = (cid:0) 𝐵𝑊

2 and 𝑘2 = 𝐵𝑊 3

22𝑆𝐹 . Its curve

function is the following:

𝑓𝑞𝑢𝑎𝑑𝑟𝑎𝑡𝑖𝑐 „𝑡” = (cid:0)

𝐵𝑊

2

‚

𝐵𝑊 3
22𝑆𝐹

𝑡2

(4.9)

By subtracting Equation 4.8 from Equation 4.9, we can obtain a frequency-shift function 𝑓Δ„𝑡” as

follows:

𝑓Δ„𝑡” =

𝐵𝑊 2
2𝑆𝐹

𝑡 (cid:0)

𝐵𝑊 3
22𝑆𝐹

𝑡2

(4.10)

85

+𝑩𝑾𝟐−𝑩𝑾𝟐𝒇𝒇𝒕𝒕Figure 4.8 The illustration of the backscatter non-linear chirp demodulation with a linear base
up-chirp as excitation.

Given the values of SF and BW, the MCU can accurately compute the frequency shift at any time.

For different non-linear types at different tags, we change the coefficients 𝑘𝑖 in Equation 4.7 to

calculate the corresponding frequency-shift function 𝑓Δ„𝑡”.

4.3.4 Non-linear Chirp Energy Concentration

As we use square wave signals to add a frequency shift on a linear excitation chirp, the backscat-

ter signals will contain two non-linear chirps that are symmetric about the linear excitation chirp.

Take a linear excitation base up-chirp as an example. As shown in Figure 4.8, a gateway will re-

ceive three signals, including a red linear excitation chirp, an orange targetted non-linear chirp and

a blue mirror non-linear chirp. In addition, a green chirp is the base non-linear down-chirp used

for demodulating the targeted non-linear chirp. With a non-linear dechirp, as shown in the right

sub-figure, the energy of the mirror non-linear and linear excitation chirps will be spread over mul-

tiple FFT bins due to the energy scattering effect of non-linear chirps. Thus, the targeted non-linear

chirp can be demodulated successfully even when the interference of the linear excitation chirp is

stronger.

However, when the linear excitation chirp is no longer a base up-chirp, the energy of the targeted

non-linear chirp will spread to two peaks on the spectrum after the non-linear dechirp, as shown

in Figure 4.9. The reason is the frequency offset between the non-linear base down-chirp and

the targeted non-linear chirp, degrading the interference tolerance ability. To achieve the same

interference tolerance ability by taking a linear base up-chirp as the excitation signal, we manually

86

𝑨𝒎𝒑𝒍𝒊𝒕𝒖𝒅𝒆+𝑩𝑾𝟐−𝑩𝑾𝟐𝒇𝟎𝟐𝑺𝑭−𝟏𝑭𝑭𝑻𝒃𝒊𝒏𝒕Figure 4.9 The illustration of the two energy peaks on the spectrum after a non-linear dechirp with
a normal linear excitation chirp.

combine the two energy peaks on the spectrum to ensure that the energy of the whole targeted

non-linear chirp is concentrated. To ensure the two peaks can be superposed coherently, we search

for all possible phase differences between the two peaks and select the phase difference that can

maximize the energy of the superposed peaks.

4.3.5 Chirp Offset Removal

In practice, compared to the linear excitation chirps, non-linear backscatter chirps from a Prism

tag experience a propagation delay as its ToF (time of flight), leading to an unexpected time offset

(TO) between the non-linear chirps and their base down-chirps during demodulation.

In addi-

tion, another misalignment may also be caused by carrier frequency offset (CFO). In the non-linear

dechirp, the CFO and TO will spread the spectrum power of non-linear chirps into multiple fre-

quency bins. Thus the backscatter symbols cannot be demodulated successfully.

The Start-Frame-Delimeter (SFD) of a linear excitation LoRa packet contains two linear base

down-chirps. After a Prism tag detects the preamble of the linear excitation LoRa packet, based on

the linear base down-chirps in the SFD, it will generate two non-linear base down-chirps, which are

to calibrate the TO and CFO. Similar to multiplying a base down-chirp to up-chirps and conducting

FFT in the dechirp, we define a reverse-dechirp as multiplying a base up-chirp to down-chirps and

conducting FFT. The basic observation is that when we apply the reverse-dechirp to the two non-

linear base down-chirps, the CFO makes the energy peaks far away from FFT bin 0 and the TO

87

𝑨𝒎𝒑𝒍𝒊𝒕𝒖𝒅𝒆+𝑩𝑾𝟐−𝑩𝑾𝟐𝑓𝑭𝑭𝑻𝒃𝒊𝒏Figure 4.10 The implementation of an excitation LoRa node, Prism tags, and a USRP gateway.

makes the energy of the base down-chirps scatter in multiple FFT bins after the reverse-dechirp.

First, we search the TO to achieve the two maximum repetitive energy peaks with the reverse-

dechirp. Then, according to the FFT bins where the energy peaks appear, we calculate and calibrate

the CFO. By doing so, we successfully remove the chirp offsets, including the TO and CFO, and

the non-linear reverse-dechirp ensures the process is inter-tag interference resilient.

4.3.6 Backscatter Data Recovery

After the chirp offsets are calibrated, we use the non-linear dechirp to demodulate the backscat-

ter non-linear chirps with a sliding window. A Prism tag uses On-Off-Keying (OOK) modulation

to embed ‘0’ or ‘1’ into a non-linear chirp symbol. Specifically, if we detect an energy peak using

the non-linear dechirp in a time window, it represents bit ‘1’. Otherwise, the bit is ‘0’. Beyond

OOK, we can be easily extended to 𝑛-Frequency Shift Keying (FSK) by adding a 𝑛-level frequency

shift on the backscatter non-linear chirps to represent different bits. The different frequency off-

sets are equivalent to different time offsets during the dechirp. Considering the energy scattering

effect of non-linear chirps with different shapes or time offsets, the inter-tag interference of 𝑛-FSK

is negligible.

4.4

Implementation

We implement Prism tags using the same hardware framework with P2LoRa [168]. The pream-

ble detection module includes a three-stage voltage-doubling amplifier HSMS-285C [174] and a

low-power voltage comparator LPV7215MG [175]. The backscatter module consists of an MCU

88

GatewayUSRP N210SBXDaughter BoardNooelec Lana LNALoRa NodeDragino ShieldArduino UNOTagDragino ShieldArduino UNOSTM32L011D3P6 [176], a low-power DAC MAX5530 [177], a low-power VCO LTC6990IS6 [178],

and a reflective radio frequency switch ADG902 [179]. The excitation source is a COTS LoRa node

with Semtech SX1276 [134] radio chip. For a gateway, we use the UHD+GNU-radio to control

a USRP N210 [152] with an SBX 400-4400 RF daughterboard [153] to record raw signals in the

air, which are further processed in MATLAB. We use a 3 dBi gain omnidirectional antenna for

Prism tags and the USRP. The total power consumption of a Prism tag is 695𝜇𝑊 during working.

Figure 4.10 shows the implementation of the Prism tag, the LoRa node, and the USRP.

4.5 Evaluation

In this section, we evaluate the performance of Prism to answer the following questions:

(cid:15) Q1 (§V-A): In outdoor environments, does Prism perform as reliably as existing LoRa backscatter

with linear chirps for long-distance backscatter communication?

(cid:15) Q2 (§V-B): In indoor environments, does Prism perform as reliably as existing LoRa backscatter

does with linear chirps for long excitation distance?

(cid:15) Q3 (§V-C): In general, can Prism reliably scale to the scenario of massive concurrent transmis-

sions?

Basic Experimental Settings: The default SF, BW, coding rate (CR), and transmission power

of the excitation LoRa node are set as 10, 125 kHz, 4

5, and 20 dBm. The SF setting is consistent

among the excitation LoRa node, backscatter tags, and gateway. For the indoor and outdoor deploy-

ments, we define the distance between the excitation Lora node and a tag as source-to-tag distance

(𝑑𝑆𝑇 ). Similarly, the distance between a tag and the gateway is defined as tag-to-receiver (𝑑𝑇 𝑅). At

each location/configuration, the excitation LoRa node transmits 300 LoRa packets, and a Prism tag

modulates 28 bits of data based on an excitation LoRa packet to calculate the bit error rate.

Performance Metrics: We use two performance metrics as follows to indicate the reliability and

scalability of a LoRa backscatter system.

• Bit Error Rate (BER) is a metric to evaluate the performance of communication reliability

over a channel that has been altered due to noise, interference, distortion, or bit synchroniza-

89

Figure 4.11 The deployment of our outdoor experiments.

tion errors. BER is widely used to demonstrate the channel noise resilience of a communi-

cation system. The lower the BER is, the higher the reliability of the system is. Similar to

P2LoRa, we set BER as 10(cid:0)4 if all data bits are successfully decoded.

• Network Throughput refers to the total data delivered through a network system in a unit of

time. Network throughput is critical to show the scalability of a network system. The higher

the throughput is, the better the scalability is. In our experiments, we use bits per second

(bps) or Kbps as the unit to quantify the network throughput.

Baseline Method: We compare Prism with state-of-the-art concurrent LoRa backscatter P2LoRa [168].

4.5.1 Outdoor Experiments

Experiment Setup: Figure 4.11 shows the deployment of our outdoor experiments. We deploy

a Prism tag at different locations (e.g., blue locations) on our campus. It uses a quadratic type of

non-linear chirps defined as 𝑓 „𝑡” = 𝑡2. The excitation LoRa node is put at the locations (e.g., green

locations) close to the tag, and 𝑑𝑆𝑇 is 5 m. We put the USRP gateway (e.g., orange location) at a

fixed location. Then, we move the pair of the excitation LoRa node and the tag. Thus, 𝑑𝑇 𝑅 varies

from 50 m to 600 m. We repeat the same experiments by setting SF as 9.

Results: As shown in Figure 4.13, we can see that the BER is almost the same between P2LoRa

and Prism at all locations. If 𝑑𝑇 𝑅 is not greater than 400 m, both Prism and P2LoRa can decode all

data bits successfully. When 𝑑𝑇 𝑅 comes to 500 m or larger, the BER gets larger. For example, the

90

LoRanodeTagGatewayLoRa nodePrismtagGatewayFigure 4.12 Outdoor BER comparison (SF=9). Figure 4.13 Outdoor BER comparison (SF=10).

Figure 4.14 The deployment of our indoor experiments.

BER is 1.5% when 𝑑𝑇 𝑅 is 500 m. When 𝑑𝑇 𝑅 increases to 600 m, the BER of Prism is 14.9%, which

is 1.5% higher than P2LoRa. Such a difference can be mitigated by upper-layer bit error correction

methods [180].

In Figure 4.12, a similar trend can be observed when we set SF as 9. Since SF-9 chirps are

less noise resilient than SF-10 chirps, when 𝑑𝑇 𝑅 is equal to or less than 300 m, both systems are

able to decode all the transmitted bits reliably. As the distance increases to 400m, the BER of both

P2LoRa and Prism increases to 2.5%. In addition, as the distance increases to 600m, the BER level

increases to 50%.

Remark: Prism realizes a similar backscatter distance to P2LoRa. Therefore, non-linear chirps can

be adopted to enable long-range LoRa backscatter.

Experiment Setup: Figure 4.14 illustrates the deployment of our indoor experiments. We deploy

Prism in our office building. The gateway is put in our office, and the excitation LoRa node and

91

LoRanode Gateway Tag47m25mFigure 4.15 Indoor BER comparison (SF=10).

Figure 4.16 Indoor BER comparison (SF=9).

tag are deployed in a corridor. A wall separates the line of sight path between the gateway and the

backscatter tag. We use the same quadratic non-linear chirp as we do in the outdoor experiments.

The 𝑑𝑇 𝑅 is fixed to 15 m, and the 𝑑𝑆𝑇 varies from 1 m to 30 m. We repeat the same experiments by

setting SF as 9.

Results: As shown in Figure 4.16, as P2LoRa does, Prism can reliably decode the data bits when

the 𝑑𝑆𝑇 is not greater than 20 m. When the 𝑑𝑆𝑇 is larger than 20 m, the BER of Prism and P2LoRa

increases since the excitation linear chirps become weaker with the large 𝑑𝑆𝑇 . In addition, Prism

achieves a bit lower BER than P2LoRa. However, the achieved BER is less than 1%, indicating

similar communication reliability.

When the SF is set to 9, we observe a similar trend in Figure 4.15. We can see that the 𝑑𝑆𝑇 is

reduced to 15 m for reliable communication since SF-9 chirps are more vulnerable to noises. The

communication reliability decreases quickly as the 𝑑𝑆𝑇 is larger than 20 m. The BER of Prism can

reach 60.3% when the 𝑑𝑆𝑇 is 30 m which is much higher than that using SF-10 chirps.

4.5.2

Indoor Experiments

Remark: Prism experiences a similar influence brought by enlarging the distance between the tag

and the excitation LoRa node to P2LoRa. Thus, non-linear chirps are a perfect alternative for linear

chirps in LoRa backscatter systems.

4.5.3 Overall Throughput under Collision

Experiment Setup: We conduct a trace-driven emulation to evaluate the network throughput of

Prism in various backscatter collisions. We collect different types of backscatter non-linear and

92

Figure 4.17 Prism uses seven types of non-linear chirps.

linear chirps in our indoor deployment. Then, we superpose these collected backscatter chirp sym-

bols to create different backscatter collisions. A total of seven different types of linear or non-linear

chirps are used in this experiment. These chirp types and the corresponding time-frequency func-

tions are listed as follows:

• 𝑙𝑖𝑛𝑒𝑎𝑟 : 𝑓 „𝑡” = 𝑡

• 𝑞𝑢𝑎𝑑𝑟𝑎𝑡𝑖𝑐1 : 𝑓 „𝑡” = 𝑡2

• 𝑞𝑢𝑎𝑑𝑟𝑎𝑡𝑖𝑐2 : 𝑓 „𝑡” = (cid:0)𝑡2 ‚ 2𝑡

• 𝑞𝑢𝑎𝑟𝑡𝑖𝑐1 : 𝑓 „𝑡” = 𝑡4

• 𝑞𝑢𝑎𝑟𝑡𝑖𝑐2 : 𝑓 „𝑡” = (cid:0)𝑡4 ‚ 4𝑡3 (cid:0) 6𝑡2 ‚ 4𝑡

• 𝑠𝑖𝑛𝑒1 : 𝑓 „𝑡” = 𝑠𝑖𝑛„𝑡”, 𝑡 2 »(cid:0)𝜋(cid:157)2, 𝜋(cid:157)2”

• 𝑠𝑖𝑛𝑒2 : 𝑓 „𝑡” = 𝑠𝑖𝑛„𝑡”, 𝑡 2 »(cid:0)3𝜋(cid:157)8, 3𝜋(cid:157)8”

The shapes of the base linear and non-linear up-chirp are shown in Figure 4.17.

First, we compare Prism with P2LoRa when the collisions happen in the same channel. Specif-

ically, Prism uses the seven different types of chirps to modulate data bits for 7 concurrent tags, but

the 7 tags only use linear chirp to modulate data in P2LoRa. We randomly select n (n (cid:20) 7) tags to

transmit data concurrently. Second, we emulate the collision scenarios across multiple channels.

We adopt the method of P2LoRa to assign different tags to 100 different channels. The number

93

Linearquadratic1quadratic2quartic1quartic2sine1sine2+𝑩𝑾𝟐−𝑩𝑾𝟐𝒇𝒕Figure 4.18 Throughput comparison between
P2LoRa and Prism in the same channel.

Figure 4.19 Throughput comparison between
P2LoRa and Prism across multiple channels.

of concurrent tags increases from 1 to 700. When the number of concurrent tags is less than 100,

we assign each tag with a different channel using linear chirps as P2LoRa does. When the number

continuously increases, we uniformly assign a newly coming tag to a random channel, then select a

type of non-linear chirp that is not used by other tags in the channel. In comparison, P2LoRa assigns

channels uniformly and randomly to new tags which use linear chirps only. In addition, to mimic

the near-far effect in the collisions, we randomly select a Signal-Interference-Ratio (SIR) between

any two collided non-linear chirp symbols in the range between -10 dB and 0 dB [171, 172]. Given

the number of collisions, we generate 40 collision scenarios with diverse types of non-linear chirps,

channel assignments, and SIR to calculate the average throughput as our results. If the data bits
of a tag are successfully decoded, the throughput of the tag equals 𝐵𝑊
2𝑆𝐹 (cid:2) 𝐶 𝑅 = 97.7 bps. Then,
the throughput of a collision scenario is calculated as the number of successfully decoded tags

multiplies the throughput of a single tag.

Results: With only one channel, as shown in Figure 4.18, the throughput of Prism linearly increases

as the number of backscatter tags increases. However, as the number of tags increases, the through-

put of P2LoRa decreases and approaches zero eventually. The results verify that the seven types of

linear and non-linear chirps are orthogonal enough to enable successful concurrent transmissions

among seven tags. Compared to P2LoRa, the backscatter concurrency is improved by 6(cid:2).

When we use 100 channels and 700 tags, as shown in Figure 4.19, the throughput of Prism

increases linearly as the number of concurrent tags increases. This verifies that Prism successfully

supports 700 tags to transmit concurrently. Prism achieves the maximum throughput 68.36 Kbps

94

1234567Number of tags0200400600800Throughput(bps)PrismP2LoRa0100200300400500600700Number of tags020406080Throughput(kbps)PrismP2LoRawith 700 tags. When the number of concurrent tags is not greater than 100, P2LoRa exhibits a

similar throughput increase as Prism. However, as the number of tags increases beyond 100, the

throughput of P2LoRa drops to zero gradually. When the number of tags reaches 700, the awful

throughput of P2LoRa can be expected due to the severe throughput loss in each signal channel.

Remark: Prism utilizes the energy scattering effect of different types of non-linear chirps to in-

crease the concurrency in each channel, thus improving the throughput significantly. In addition,

Prism is parallel to P2LoRa and can be easily extended to scenarios with multiple channels. More-

over, we only demonstrate seven types of non-linear chirps. Hopefully, we can further improve the

throughput of Prism by involving new types of non-linear chirps. How to achieve the maximum

concurrency with non-linear chirps is still an open question.

4.6 Related Work

Generally, a backscatter system uses a radio frequency switch to either reflect or absorb exci-

tation signals. The batteryless property makes it attractive in many IoT applications. For example,

Liu et al. [181] propose ambient backscatter that uses some ubiquitous wireless signals (e.g., TV

and cellular) in surrounding environments to obtain energy for backscatter communication. Zhao et

al. [182] propose OFDMA-enabled Wi-Fi backscatter that enables tags to produce backscatter sig-

nals at the frequencies of orthogonal subcarriers, which are subsequently synthesized into an OFDM

symbol by a receiver. Prism focuses on LoRa backscatter and achieves a scalable backscatter system

with interference-tolerant concurrent transmissions enabled by different types of non-linear chirps.

We summarize the related work from the following aspects.

LoRa Backscatter: Talla et al. [167] propose LoRa backscatter that uses a single-tone excitation

signal to generate a LoRa packet by shifting the frequency with a radio frequency switch. With a

specific excitation signal, Netscatter [169] further proposes to combine CSS and OOK modulation,

assigning different cyclic shifts of a linear chirp to different concurrent tags, to support parallel

decoding of concurrent LoRa packets. By taking COTS LoRa linear chirps as excitation signals,

PLoRa [166] adopts an FPGA to shift LoRa linear chirps by two frequency offsets to modulate data.

Moreover, P2LoRa [168] develops a modulation method to support parallel decoding with ambient

95

LoRa linear chirps as excitation signals. However, the scalability of P2LoRa relates to the available

bandwidth recourses, which are usually limited in LoRa, a narrow-band communication technol-

ogy. Additionally, P2LoRa suffers from inter-tag inference, which may comprise its performance

in practice. In comparison, Prism is parallel with existing methods to enhance the scalability of

LoRa backscatter systems by exploring the quasi-orthogonal logical channels created by different

types of non-linear chirps in the same physical channel.

Non-linear Chirps: Non-linear frequency modulation (NLFM) [183] has been widely used in radar

systems, which is a cost-effective pulse compression method [184] with high resolution, improved

SNR, and effective interference suppression. NLFM modulation function has an inherent spectrum

weighting that results in lower sidelobe levels compared with linear frequency modulation(LFM).

In LoRa, CurvingLoRa [171] adopts non-linear chirps to improve the scalability of LoRa networks

by enabling a LoRa gateway to successfully decode concurrent transmissions with the same type of

non-linear chirps. Moreover, CurveALOHA [172] proposes a media access control (MAC) protocol

with multiple types of non-linear chirps to increase the network throughput. In contrast, Prism

is a scalable LoRa backscatter system, which converts ambient linear LoRa chirps emitted from

COTS LoRa radios to multiple non-linear counterparts, which can be reliably decoded at a gateway

simultaneously.

LoRa Collision Resolving: LoRaWAN [1] adopts the least restrictive MAC protocol ALOHA [185],

which enables nodes to transmit data upon waking up. When multiple LoRa packets from different

nodes are transmitted concurrently, collisions inevitably appear. Existing works focus on extract-

ing distinguishable features of the collided packets in the time or frequency domain. For example,

Choir [186] matches data bits to a LoRa node with its frequency change due to its oscillator defi-

ciency. FTrack [187] divides collisions by exploring the distinct tracks on the spectrum and symbol

edges on the time domain. CIC [188] combines the spectrum from different parts of a chirp symbol,

proposing a sub-symbol method to cancel interference signals under ultra-low SNR environments.

mLoRa [189] utilizes the consecutive interference cancellation to iteratively demodulate the data

bits with no interference. CoLoRa [190] utilizes the spectral peak ratio of a misaligned chirp as a

96

characteristic to differentiate between overlapped LoRa packets. Pyramid [191] applies a sliding

demodulation window to track the changes in energy peaks to resolve collisions. NScale [192] im-

proves noise resistance in the iterative peak recovery by varying peak scaling factors within two

consecutive windows. PCube [193] classifies chirp symbols from concurrent packets by utilizing

unique phases of the air channel. However, since backscatter tags usually have no back-off mecha-

nism, the time offset among the collided chirp symbols is almost zero, which is hard to be handled

by the existing methods. In addition, the existing methods suffer from the near-far problem, in-

dicating a situation that a strong signal from a near node significantly reduces the SNR of a weak

signal from a far node. Moreover, LMAC [194] and p-CARMA [195] operate at the MAC layer and

try to use Channel Activity Detection (CAD) and CSMA (carrier-sense multiple access) to detect a

potential collision and avoid it with a backoff. However, the CAD and CSMA are energy-heavy for

a LoRa backscatter system. In contrast, Prism adopts non-linear chirps, creating new logic channels

to avoid these issues for practical LoRa backscatter concurrent transmissions.

4.7 Conclusion

To conclude, we develop a high-throughput LoRa backscatter system Prism by using the differ-

ent types of non-linear chirps to enable concurrent transmissions of multiple backscatter tags in the

same channel. Specifically, Prism takes ambient LoRa linear chirps from COTS LoRa nodes as the

excitation source. Then, a Prism tag uses a timer to trigger a low-power frequency shifting process,

which transforms the linear excitation chirps to the non-linear backscatter chirps according to an ac-

curate frequency-shift function. At the gateway side, we concentrate the energy of non-linear chirps

during the dechirp and decode the data bits encoded by backscatter tags. We prototype Prism tags

and gateways with customized PCB circuits and USRP, respectively. We evaluate Prism’s perfor-

mance in indoor and outdoor environments. Prism achieves the highest throughput of 68.36 Kbps

and supports 700 tags to transmit concurrently with a bit error rate that is less than 1%. With more

non-linear chirp types, Prism is expected to perform better.

97

CHAPTER 5

AEROECHO: TOWARDS AGRICULTURAL LOW-POWER WIDE-AREA
BACKSCATTER WITH AERIAL EXCITATION SOURCE

5.1

Introduction

Recently, the Internet of Things (IoT) has revolutionized agricultural methods, giving rise to

smart agriculture [196]. This breakthrough incorporates interconnected agricultural sensors, pro-

foundly boosting efficiency, productivity, and sustainability in managing both crops and livestock [2,

15,18,197–199]. Crucial to this transformation is the demand for cost-effective hardware with broad

communication range capabilities, vital for spanning vast agricultural landscapes [11, 200, 201].

With the deployment of numerous sensors, low power requirements are essential, consequently

reducing the frequency of battery replacements and maintenance [5, 196, 202].

Backscatter communication, a cutting-edge IoT technology, utilizes existing carrier waves to

modulate information bits, eliminating the need for amplifiers and carrier modulation found in

active radio systems [203]. Recent research on Long Range (LoRa) backscatter [166–169] aims to

integrate LoRa communication with backscatter radios, achieving low energy consumption (e.g.,

(cid:20)1 mW) while sustaining long communication distances between tags and receivers (e.g., 1.1km

to 2.8km), potentially meeting the demands of agricultural IoT: long-distance, low-power, and cost-

effectiveness [7].

However, several factors currently make long-range backscatter communication impractical, as

summarized in Table 5.1. The first issue is the high deployment cost. The effective communication

range between the excitation source and the backscatter tag in existing solutions is roughly 20(cid:2)

shorter than the distance between the tag and a LoRa gateway, spanning less than 15 meters in a

single tag static backscatter system [1, 166, 167, 204]. This restriction forces the deployed position

of tags to be very close to the excitation source compared to the distance to the receivers, leading to

the dense deployment of excitation sources and increased costs. Additionally, the dense tag deploy-

ment causes collisions and reduces scalability [1]. Some existing works do not support multi-tag

backscattering [166, 167, 204], while others require high spectrum occupation to support multi-tag

98

Figure 5.1 Illustration of AeroEcho.

networking [168, 169, 205]. This limitation prevents the scalability of tags in agricultural settings,

restricting the potential for high throughput. Moreover, long excitation systems [206–208] focus

solely on shortening the distance between the excitation source and the tag and reducing costs, but

they overlook the importance of scalability.

This paper introduces AeroEcho, a novel low-power wide-area data collection system that uti-

lizes backscatter with a high-speed aerial excitation source and fixed gateways. The concept is

illustrated in Figure 5.1. Inspired by the wide range of usage of drones as infrastructure in modern

smart farms [209], AeroEcho shifts the excitation source from stationary locations to unmanned

aerial vehicles (UAV), reducing the cost of dense excitation source deployment for tags widely dis-

tributed in the farmland. The excitation source and tag are co-designed for multi-tag decoding in

the same frequency channel to boost scalability. We deploy receivers with fixed gateways because

of the complexity of full-duplex transceivers and limited computation and storage resources on the

drones. However, achieving AeroEcho design entails overcoming three significant challenges:

The initial challenge is that when one excitation source signal is transmitted, all the backscatter

tags within the communication range can be woken up and start backscatter modulation due to the

standard synchronized symbols in the commercial-off-the-shelf (COTS) excitation device. This

inevitably leads to signal collision and requires excitation signal cancellation at the receiver side.

To address this challenge, we co-design our excitation source and tag with a customized packet

format to avoid unnecessary synchronization and collision for backscatter signals. Furthermore,

we propose asynchronous decoding methods to distinguish and decode the signals from different

99

Table 5.1 Existing long-range wide-area backscatter comparison. (Tag data rate is measured in
SF12 and 125kHz bandwidth)

Single tag [166, 167, 204]
Parallel
decoding [168, 169, 205]
Long excitation [206–208]
AeroEcho

Cost

High

High

Low
Low

Scalability

Low

Low

Low
High

Max
Throughput
13.6kbps

250kbps

19.6kbps
1.46Mbps

Tag data rate

24.5bps

30.5bps

-
366bps

tags without the complex carrier cancellation process at gateways.

Next, the random and dense distribution of tags on the land poses a challenge. More tags with

asynchronous decoding can decrease communication performance due to a higher symbol error

rate (SER). Achieving an optimal balance between the number of active tags and SER to maximize

throughput becomes imperative. To tackle this issue, we introduce an excitation cell methodology

to facilitate optimal multi-tags decoding performance. Specifically, the UAV selectively activates

only those tags within a predefined circular area designated as the excitation cell. The radius of this

cell is calibrated to encompass the maximum permissible distance from the source to any given tag.

Lastly, accounting for UAV range constraints, tag energy efficiency, and unpredictable tag place-

ment, designing a flight path that ensures comprehensive coverage of multiple radio excitation cells

within a specific area presents a formidable challenge. This challenge is formalized as an optimiza-

tion problem, emphasizing either the energy efficiency of backscatter tags or the range efficiency

of UAVs while considering constraints related to communication and coverage reliability. To tackle

this issue, we have devised two distinct route planning strategies: Rectangular Displacement and

Annular Trajectory, tailored to optimize coverage and energy efficiency.

We implement AeroEcho utilizing customized PCB circuits and software-defined radios op-

erating on TV white space spectrum, thereby circumventing the 0.4-second on-air time limita-

tions inherent in industrial, scientific, and medical (ISM) bands and extending communication dis-

tances [210]. Our evaluation encompasses real-world farm scenarios and trace-driven large-scale

emulation. The results demonstrate that signals from 71 tags can be decoded with less than 1% SER

100

on the same frequency channel, showcasing a 10(cid:2) increase in concurrency compared to existing

methods. Furthermore, the overall throughput can be expanded to 1.46 Mbps across multiple fre-

quency channels. Our routing scheme yields up to 1.5(cid:2) lower energy consumption for tags and up

to 9(cid:2) shorter range for UAVs. In summary, the contributions of this paper are delineated as follows:

• We propose a comprehensive LoRa backscatter system with an aerial excitation source to

enable scalable agricultural IoT.

• We develop practical methodology and efficient algorithms to optimize backscatter concur-

rency, network throughput, and energy efficiency, simultaneously.

• We prototype AeroEcho and evaluate its performance in real environments, emulation, and

simulation. The results show that the maximum concurrency of AeroEcho is 10(cid:2) of the

state-of-the-art with the same frequency resources. Moreover, AeroEcho improves the overall

throughput by 5.84(cid:2) and individual tag data rate by 12(cid:2).

5.2 Preliminary

5.2.1 Backscatter in Agricultural IoT

IoT devices predominantly utilize battery-powered sensors for data transmission [11]. The

transmission process involves generating baseband signals, modulating these signals with carrier

signals, and amplifying the signals. The consumption of substantial energy for carrier modulation

and amplification presents an issue for devices requiring frequent agricultural sensor data report-

ing. In contrast to active radios, backscatter radios use signals from excitation sources as carrier

signals and modulate their data on top of carrier signals, significantly reducing power consump-

tion [166, 203].

Typically, a backscatter system includes an excitation source, a tag, and a receiver. We can shift

power consumption from the sensor node (tag) to a shared ambient source, where the energy use

for wireless communication is negligible. However, rural farmlands lack infrastructure. Although

low-power wide-area networks (LPWANs) are available, their gateways are sparsely distributed

101

kilometers apart [7, 10], unable to support long excitation distances or many concurrent tag trans-

missions. Instead, drones, widely used in fertilization and irrigation on smart farms, are a good

choice for bringing the source close to the tag [2, 209].

5.2.2 LoRa Backscatter

Recent years have seen significant advancements in the design of LoRa-based backscatter ra-

dios, combining the low-power backscatter technology with long-range techniques to improve per-

formance [166, 167]. LoRa is designed for IoT up to 10 km with Chirp Spread Spectrum (CSS)

modulation [4, 154]. The basic unit of LoRa modulation is a linear up-chirp whose frequency in-

creases linearly with time across the whole bandwidth [40, 211]. The key to LoRa modulation is

that a time delay in a chirp can be transformed into a cyclic frequency shift. The initial frequency

can modulate encoded data bits as cyclic time shifts. The demodulation process is ‘dechirp’ as de-

fined in Equation 5.1. Where 𝑓0 represents the initial frequency and 𝑓𝑐 is the up-chirp. It multiplies

a received chirp symbol with a base down-chirp (the conjugate of the base up-chirp (cid:0) 𝑓𝑐 „𝑡”) whose

frequency decreases linearly over time. After the Fast Fourier Transform (FFT), a peak appears at

an FFT frequency bin, corresponding to the initial frequency of the received chirp symbol. LoRa

defines N different initial frequency offsets to encode 𝑙𝑜𝑔2𝑁 bits.

(

𝑒 𝑗2𝜋„ 𝑓0‚ 𝑓𝑐 „𝑡””𝑡

)

(

(cid:3)

𝑒(cid:0) 𝑗2𝜋 𝑓𝑐 „𝑡”𝑡

)

= 𝑒 𝑗2𝜋 𝑓0𝑡

(5.1)

5.2.3 Asynchronous Decoding Rationale

When multiple tags are excited by the same COTS excitation source, traditional linear chirp-

based backscatter signals will suffer severe collision issues. We can observe multiple energy peaks

on the spectrum and cannot distinguish signals reliably [1, 205]. Netscatter and P2LoRa [168, 169]

solve the problem by taking multiple frequency channels.

The nature of a non-linear chirp [171] makes it easy to solve collision issues. It still utilizes CSS

modulation, replaces the linear chirp signals 𝑓𝑐 „𝑡” with a non-linear chirp time-frequency function,

and the dechirp of non-linear chirps can be expressed in Equation 5.1.

Orthogonality among different non-linear chirp types. The different types of non-linear chirps [205]

102

Figure 5.2 Asynchronous decoding rationale.

define the math function of non-linear quadratic chirp as 𝑓𝑛𝑜𝑛1„𝑡” = 𝑘1𝑡2 ‚ 𝑘2𝑡 ‚ 𝑘0 and quartic func-

tion as 𝑓𝑛𝑜𝑛2„𝑡” = 𝑧1𝑡4 ‚ 𝑧2𝑡3 ‚ 𝑧3𝑡2 ‚ 𝑧4𝑡 ‚ 𝑧0. After demodulation with non-linear chirp, the output

signals are shown as follows respectively:






𝐹𝑛𝑜𝑛1„𝑡”

= 𝑓0,

𝐹𝑙𝑖𝑛𝑒𝑎𝑟 „𝑡” = 𝑓0 ‚

∑
1
𝑖=0

∑

𝑥𝑖𝑡𝑖 (cid:0)

𝐹𝑛𝑜𝑛2„𝑡”

= 𝑓0 ‚

∑
4
𝑚=0

𝑧4(cid:0)𝑚𝑡𝑚 (cid:0)

2
𝑗=0

𝑘 𝑗 𝑡 𝑗 ,
∑
2
𝑛=0

𝑘𝑛𝑡𝑛

(5.2)

Only 𝐹𝑛𝑜𝑛1„𝑡” is one constant. A corresponding peak can be detected on the spectrum, while

the energy of mismatched type symbols spreads over the spectrum. However, the orthogonal non-

linear chirp types are limited and complicated to design and implement. There are only 6 given

non-linear chirp formulas [171, 205]. It is hard to meet the requirements of hundreds of sensors

deployed in agriculture.

Asynchronous decoding among the same non-linear chirp type. Further, an intrinsic property

of non-linear chirps necessitates precise temporal alignment. This requirement provokes a recon-

sideration of our approach to non-linear chirp backscatter. The frequency function of non-linear

is time-variant, leading to the spectrum energy distribution changing over time. As shown in Fig-

ure 5.2, the energy of red interference symbols spreads over multiple, clustered FFT bins due to the

misalignment with the dechirp window. If we demodulate a quadratic non-linear chirp with a time

offset 𝑡𝑔𝑎 𝑝, the mathematical function of dechirp can be expressed as follows:

𝐹𝑛𝑜𝑛1(cid:0)𝑜 𝑓 𝑓 𝑠𝑒𝑡 „𝑡” = 𝑓0 ‚ 𝑘2𝑡2

𝑔𝑎 𝑝 ‚ 2𝑘1𝑡𝑔𝑎 𝑝 (cid:2) 𝑡 ‚ 𝑘2𝑡𝑔𝑎 𝑝

(5.3)

103

+𝐁𝐖𝟐−𝐁𝐖𝟐Frequency𝐭Amplitude𝟎𝐅𝐅𝐓𝐛𝐢𝐧𝑡𝑔𝑎𝑝Figure 5.3 System overview.

The output frequency is a quadratic function of time instead of a constant. The scattering effects

make it easy for the target symbol (blue solid line) to be demodulated, and we can see a strong

energy peak on the spectrum. As a result, the energy of the interference symbol can be regarded

as noise. We can do asynchronous decoding by sliding the dechirp window. The same type of

non-linear chirps builds multiple quasi-orthogonal logical channels by time offsets among different

chirps. However, creating time offsets cannot be applied to backscatter systems with standard LoRa

excitation sources. This motivates us to co-design excitation-tag.

5.3 System Design

As shown in Figure 5.3, AeroEcho consists of a mobile excitation source on a UAV, backscatter

tags, and a gateway. The basic unit of AeroEcho is the excitation cell. The single excitation cell

takes the UAV excitation signals transmission location as center and the maximal excitation-source-

to-tag distance as radius. There are multiple tags distributed in a single cell. Tags can be woken up

(5.3.1) and modulate their own data on top of the carrier signals with a random time delay (5.3.2).

At gateways, AeroEcho removes offsets and demodulates symbols (5.3.3). RF source on UAV

transmits excitation signals with a prescribed coverage scheme (5.3.4) across multiple excitation

cells. Finally, we can recover data bits from various tags.

104

AerialSourceGatewayTime & FreqOffsetRemovalAsync DecodePackets from multi-tagsAgriculturalDataRandomDelayCustomizedPreambleExcitation-TagCo-designModulationAeroEchoExcitationCellAerialCoverage SchemeTVWS Spectrum(a) AeroEcho excitation signals

(b) AeroEcho backscattered signals

Figure 5.4 Illustration of AeroEcho excitation source signals and backscattered signals.

5.3.1 Excitation Source and Tag Co-design

When a UAV reaches the excitation point, it transmits preamble signals to activate AeroEcho

backscatter tags. Initially, we employ eight repeated linear chirps as preamble signals, a design

choice that aligns with the standard LoRa. We devised a passive preamble detection circuit to iden-

tify the arrival of these excitation signals. This circuit comprises impedance matching, an envelope

detector, and a comparator. The impedance matching and the envelope detector are designed to

recognize the repetitive pattern of the repeated linear chirps in the preamble. The comparator, the

third component, evaluates if our tag can be activated by juxtaposing the reference voltage with

the output voltage from the first two components. The excitation signals format is illustrated in

Figure 5.4a. We adjust the amount and amplitude of linear chirps in the preamble with differ-

ent transmission power to change the sensitivity of the preamble detection circuit. This enables

AeroEcho tags to achieve different excitation source-to-tag distances in different excitation cells

under various environmental factors (e.g. weather, agricultural activities) without additional hard-

ware or software modification. Adjusting a UAV instead of individually reconfiguring multiple tags

makes the system adaptive and scalable. Then, we utilize single-tone signals as carrier signals to

allow flexible modulation for non-linear chirps.

As Section 5.2.3 mentions, time offsets among different chirp symbols can lead to scattering

effects to distinguish signals from other tags. When a UAV hovers at a specific location and emits

excitation signals, it can cover the backscatter tags within the circular area defined by its excitation

cell. However, in agricultural settings, tags equipped with sensors are often distributed irregu-

larly [2]. On the other hand, using such circular coverage patterns cannot seamlessly cover a farm

105

…+𝑩𝑾𝟐−𝑩𝑾𝟐𝒕N linear chirpsSingle tonePreamble𝐟…+𝑩𝑾𝟐−𝑩𝑾𝟐𝐟𝒕Random DelaySFDPayloadPreambleN linear chirpswithout overlaps. If fixed offsets are assigned, it becomes challenging to synchronize sensors at

random locations with varying backscatter modulation times across multiple excitation cells. Ev-

ery time we adjust the deployment of the farm or add new sensors, we need to reschedule the delay

design among hundreds of tags. The maintenance cost is high. To enable concurrent transmission,

AeroEcho tags give different random offsets ranging from 0 to 1 symbol time as waiting time after

the preamble, as shown in Figure 5.4b. Different random time offsets among multiple tags can

create orthogonal logical channels to support concurrent transmission.

5.3.2 AeroEcho Packet Format

After preamble detection, AeroEcho tag can wake up and assign a specific frequency shift, con-

verting single-tone signals into non-linear chirps. We use a microcontroller to control the voltage

output of the digital-analog converter. The voltage is the input of a voltage-controlled oscillator.

After that, the RF switch and antenna adjust the impedance and radiate the signals, adding the

frequency shift on the carrier signals. We formulate the voltage function accordingly to create a

non-linear chirp generation that matches the signals’ desired final time-frequency shape. In the

time-frequency domain, non-linear chirp functions can be expressed as polynomial functions.

𝑓𝑐 „𝑡” =

𝑛∑

𝑖=0

𝑘𝑖𝑡𝑖, 𝑡 2 »0, 2𝑆𝐹
𝐵𝑊

…, 𝑓𝑐 „𝑡” 2 »(cid:0)

𝐵𝑊

2

,

𝐵𝑊

2

…

(5.4)

For a non-linear base up chirp of quadratic function, 𝑘0 = (cid:0) 𝐵𝑊
𝑘0 = (cid:0) 𝐵𝑊

2 , 𝑘2 = 𝐵𝑊 3
22𝑆𝐹 and for encoded chirps,
22𝑆𝐹 (mentioned in Section 5.2.3). After excitation signals iden-
tification and random time delay, we modulate single tone signals to generate two non-linear base

2𝑆𝐹 (cid:0)1 , 𝑘2 = (cid:0) 𝐵𝑊 3

2 , 𝑘1 = (cid:0) 𝐵𝑊 2

down-chirps SFD (red curves) as shown in Figure 5.4b to synchronize initial time and frequency.
This process is discussed in Section 5.3.3. For SFD chirps, 𝑘 𝑆𝐹 𝐷0 = 𝐵𝑊

2 , 𝑘 𝑆𝐹 𝐷2 = (cid:0) 𝐵𝑊 3
22𝑆𝐹 .

Like linear chirp modulation, we transfer the cyclic time offset to the initial frequency offset

for each symbol. We can generate different time-variant voltage curves with different cyclic time

offsets to modulate multiple initial frequency symbols. Then, we can implement non-linear CSS

modulation on backscatter tags, which is a considerable advantage compared to existing work with

OOK. The spreading factor can vary from 1 to 12, which means AeroEcho encodes 1 bit to 12

106

(a) Demodulation Window 1.

(b) Demodulation Window 2.

Figure 5.5 Illustration of the sliding window for non-linear chirps demodulation at the gateway.

bits for each symbol. The multiple spread factor and flexible initial frequency offset choices meet

the multiple data rate requirements. We use the Heaviside step function [212] to express the time-

frequency function of non-linear chirp with the impact of cyclic time offset (denoted as 𝑡𝑜) on the t

can be expressed as follows:

𝑓𝑛𝑜𝑛1„𝑡” = (cid:0)

𝐵𝑊

2

‚

𝐵𝑊 3
22𝑆𝐹

𝑜…
»𝑡2 ‚ „2𝐻𝑒𝑎𝑣𝑖𝑠𝑖𝑑𝑒„𝑡 (cid:0) 𝑡𝑜” (cid:0) 1”𝑡2

(5.5)

5.3.3 Asynchronous Decoding

Thanks to the single-tone excitation signals, AeroEcho do not need to operate complicated ex-

citation carrier signals cancellation. The demodulation necessitates precise time synchronization

owing to the sensitivity of non-linear chirps to time offsets, as outlined in Section 5.2.3. Random set

delays and sampling time offset induced time offsets (TO) can distribute the spectral power across

multiple frequency bins, and hardware-induced carrier frequency offset (CFO) shifts the energy

peak. According to Equation5.3 in Section 5.2.3, the initial frequency offset caused by TO and

CFO for non1 can be expressed as: Δ 𝑓 = 𝑘1𝑇𝑂2 ‚ 𝐶𝐹𝑂.

To remove offsets, after the preamble and random time delay, we generate two non-linear base

down-chirps as the start of frame delimiter (SFD) shown in Figure 5.4b. We multiply base up-chirps

with two SFD chirps and identify the TO corresponding to the spectrum’s two most substantial

repetitive energy peaks. Then, we calculate the CFO according to the location of the shifted energy

peaks. This effectively mitigates the influence of TO and CFO. Moreover, the non-linear SFD

provides resilience against interference among tags.

After offset removal, we can receive signals like Figure 5.5 shown. For example, Orange, pink,

107

+𝐁𝐖𝟐−𝐁𝐖𝟐𝑓𝑡FFT. Abs0𝐹𝐹𝑇𝑏𝑖𝑛Demodulation Window+𝐁𝐖𝟐−𝐁𝐖𝟐𝑓𝑡FFT. Abs0𝐹𝐹𝑇𝑏𝑖𝑛Demodulation Windowgrey, green, and blue symbols come from four tags. We use sliding windows of one symbol length to

demodulate signals from each tag. The sliding window aligns with the first orange chirp of initial
frequency (cid:0) 𝐵𝑊

2 , and a peak appears at the initial FFT bin. The receiver then takes other color

symbols not aligned with the demodulation window as noise. When the sliding window aligns

with the second orange symbol with an initial frequency of 0, a peak appears in the middle of the

spectrum. Similarly, we can also decode all data from different tags.

5.3.4 Aerial Coverage Scheme

The RF source at UAVs transmits excitation signals at the centers of circles with a radius of r equal

to maximal 𝐷 𝑠𝑡, creating basic excitation cells. Tags are distributed in a specific area with random

locations. We aim to plan multiple excitation cells collecting data from tags with a specific density in

a given area and keep the low SER. Meanwhile, we need to minimize the total energy consumption

of all the backscatter tags and UAV flight ranges. The problem can be defined as follows:

minimize
𝑟,𝑠,𝑁

𝐸, 𝐹 =

𝑅
𝑠

subject to SERN (cid:20) SERthresold 𝑁 (cid:21) 𝜌 (cid:1) 𝑠

(5.6)

R represents the flight range to cover the area s. E is the total energy consumption of backscatter
tags and F= 𝑅

𝑠 is the flight range efficiency of UAV, representing range per unit area. With concur-

rent backscatter transmission amount N for a single excitation cell, we must consider the symbol

error rate SERN, which is required to be less than or equal to a maximum acceptable threshold

SERthreshold. Within this area, backscatter tags are distributed randomly but with a uniform deploy-

ment density symbolized as 𝜌. The concurrency N should also be larger than the tag amount (the

product of 𝜌 and s).

We propose two methods: Rectangular displacement and Annular trajectory scheme. The rect-

angular method achieves seamless coverage but wastes energy and exhibits data collection unfair-

ness due to repetitive tag wake-ups. The annular method is energy-efficient but requires multiple

rounds to cover all tags and is not suitable for time-sensitive applications.

Rectangular Displacement Coverage: As illustrated in Figure 5.6, the gateway is at the center of

108

Figure 5.6 Rectangular Coverage Scheme

(a) First round

(b) Second round

(c) Third round

(d) Adaptive radius

Figure 5.7 Illustration of annular coverage schemes.

the figure. We have m2 excitation cells, and the column and row are m for the square deployment.

Four cells have only one intersection point and four identical overlapped orange zones (orange

shape). Then the UAV can follow the blue trajectory to travel all the excitation points and seamlessly

p

2𝑟. We can calculate the overlapped

collect all the sensor data. The width of the orange shape is

area with a geometric method as follows:





𝑆𝑜𝑣𝑒𝑟𝑙𝑎 𝑝 = 𝑚„𝑚 (cid:0) 1” „𝜋 (cid:0) 2”𝑟 2

𝐸𝑠 = 𝑒𝜌„𝑆𝑜𝑣𝑒𝑟𝑙𝑎 𝑝 ‚ 𝑆” 𝑅𝑠 = „𝑖2 (cid:0) 1”𝑟

(5.7)

However, the tags in the overlapped orange areas are triggered to perform backscatter modulation

multiple times in the same collection round. This may lead to energy waste and unfairness in IoT

sensor data collection. This creates complicated maintenance problems and is not acceptable for

some energy-sensitive applications.

Annular Trajectory Coverage: As depicted in Figure 5.7a, the dark blue cell is the initial excitation

annulus. UAV initiates the transmission of the first excitation signals from the central point of the

dark blue cell, which also serves as the gateway location. The drone then moves to the next adjacent

109

𝟐𝒓Excitation CellOverlapped ZoneAFirst Round CellBAαSecond Round CellThird Round Cellannulus, traversing the centers of the light blue excitation cells from the starting point A. When the

drone reaches all the red points and transmits the excitation signals, it is termed one round. To

cover all the tags in this annulus, AeroEcho utilizes multiple rounds at each annulus. Figure 5.7b

illustrates the second round. The shallow orange circles are excitation cells in the second round

with beginning point B. After two rounds, the angle offset on flight trajectory between A and B is

𝛼1. Above 95% tags’ data can be collected. The drone flies and transmits signals along the green

center points for each round. Likewise, in the latter rounds, the drone starts with a predefined angle
offset 𝛼 from the start point of the last round. One of the optimal settings of initial 𝛼1 is 𝜋

6 . For
the latter two rounds, we take two quartiles in »0, 𝛼𝑥… as 𝛼𝑥‚1 and 𝛼𝑥‚2. Figure 5.7c illustrates the
third round coverage scheme with the green shallow circles (𝛼 = 𝜋

12), which eventually cover over

98.7% area for the annulus.

As Equation 5.8 shows, 𝐶𝑖 symbolizes the ratio of the annular area to the average coverage area

of each round. This implies that ensuring complete coverage of all the tags within one annulus

requires a flight distance 𝐶𝑖 times the single annular trajectory’s distance. 1 (cid:20) 𝐶𝑖 (cid:20) 1.35 can

guarantee the data collection of more than 75% sensors for each round. Every tag can only be

triggered 1

𝐶𝑖 in each round on average. The experimental results of coverage rate are discussed in

Section 5.5. The flight range of UAV covering all the tags of the current annulus can be donated as

𝑅𝑐 on average.

∑

𝑑 =

𝑟 𝑗 𝐶𝑖 = 4𝑑𝑟 𝑗 (cid:1) 𝑎𝑟𝑐𝑠𝑖𝑛

𝑗

𝑟 𝑗
𝑑

∑

𝜋𝑟 2 𝑅𝑐 = 2𝜋

𝐶𝑖 (cid:1) 𝑑

𝑖

(5.8)

Given the limited battery capacity, we aim for the UAV to travel as short as possible to ensure

minimal energy consumption by backscatter tags. A larger excitation radius can expand the single-

cell area, resulting in a shorter displacement of the drone. However, a larger radius also implies

a reduced signal strength from the tag to the receiver and a shorter range. To cover with higher

efficiency, we propose an adaptive radius scheme. As depicted in Figure 5.7d, we use three excita-

tion cells with different radii. The same excitation cell is arranged in the same annulus. From the

inside out, we select the maximum radius at varying distances while ensuring the tag-to-receiver

110

Figure 5.8 Implementation of the excitation source, tags, and gateway in the farm.

distance does not exceed the current distance. This strategy provides the UAV’s range efficiency

and coverage.

5.4

Implementation

Figure 5.8 shows the devices we use for our experiments. The wake-up module consists of

a three-stage voltage-doubling amplifier HSMS-285C [174] and a low-power voltage compara-

tor LPV7215MG [175]. The modulation module includes STM32L011 [176] MCU, a low-power

voltage-controlled oscillator LTC6990IS6 [178] and a reflective RF switch ADG902 [179]. The

excitaion source is HackRF One [213] on DJI Spark at TVWS spectrum. We use GNU-radio to

control a USRP N210 [152] as a gateway, and then we do signal processing in MATLAB. The total

energy consumption AeroEcho tag is 538𝜇𝑊. The cost is less than 10 US dollars.

5.5 Evaluation

In this section, we conduct experiments to verify the performance of single tag, concurrent

transmission, throughput, data rates and UAV routing scheme. The default SF=12, bandwidth (BW)

is 250kHz, the frequency band is 470MHz at TVWS spectrum, the coding rate is 4

5, and transmission
power is 14𝑑𝐵𝑚. The default non-linear chirp type is quadratic1— 𝑓 „𝑡” = 𝑡2 as mentioned in 5.3.2.

AeroEcho tag modulates 28 symbols of information on each packet. We conduct experiments in

different scenarios with multiple source-to-tag distance(𝐷 𝑠𝑡), source-to-receiver distance(𝐷𝑡𝑟). The

SER is set to 10(cid:0)4 if all the symbols are decoded successfully.

Metrics: Symbol Error Rate (SER): SER is the ratio of data symbols being incorrectly decoded

due to noise or other impairments. A lower SER indicates a more robust and efficient transmission

111

HackRF OneLaNA LNAUSRP N210AeroEcho TagDJI Drone(a) Dst=4 m

(c) Dst=12 m
Figure 5.9 Impact of excitation cell radius to BER of AeroEcho with different 𝐷tr.

(b) Dst=8 m

system. Concurrency: The maximal tags amount to a backscatter to support simultaneously with

the corresponding SER threshold. Throughput: Actual speed at which data is successfully trans-

ferred for the backscatter networks with all the working tags. Data rates: The maximum number

of data bits transmitted for each tag per second. Energy Consumption: The total trigger times of

backscatter tags per unit area. Range Eﬀiciency: The average flight range of the UAV for covering

each tag.

Baseline: Netscatter [169], P2LoRa [168] and Prism [205].

5.5.1 Excitation-tag co-design performance

Experiment Settings: In outdoor experiments, we adopt a single tag to verify the non-linear chirp

backscatter signal performance with TVWS. We put the USRP receiver at a fixed location and

moved the tag from 50m to 350m while keeping the excitation source on the drone at 3m height

with multiple 𝐷 𝑠𝑡 =4m, 8m, or 12m in horizontal distance to AeroEcho tag. Prism and P2LoRa

with SF12 and 250kHz at 915 ISM bands are baseline methods.

Results: As shown in Figure 5.9a with Logarithm scale, we can discover that AeroEcho, Prism

and P2LoRa all can decode all the bits successfully if 𝐷𝑡𝑟 is equal to or less than 150m. As 𝐷𝑡𝑟

grows from 200m, the SER of the two baseline methods increases more than AeroEcho. When

the 𝐷𝑡𝑟 = 300𝑚, the SER of the two baseline methods reduces to about 1% while AeroEcho only

is 0.15%. Finally, the SER declines and achieves 0.009, 0.06, and 0.05 for AeroEcho, P2LoRa

and Prism, respectively. In Figure 5.9b, we can also see that the error demodulation occurs when

𝐷𝑡𝑟 = 100𝑚 for P2LoRa and Prism. When the 𝐷𝑡𝑟 = 200𝑚, the SER of AeroEcho reaches about 1%

and the SER of Prism and P2LoRa is 6(cid:2) and 5(cid:2) of the SER of AeroEcho. Finally, the SER for two

112

50100150200250300350Dtr(m)10-410-310-210-1SERAeroEchoP2LoRaPrism50100150200250300350Dtr(m)10-410-310-210-1SERAeroEchoP2LoRaPrism50100150200250300350Dtr(m)10-410-310-210-1SERAeroEchoP2LoRaPrismTable 5.2 SIR and max 𝐷𝑡𝑟 with different excitation cell radius

Radius (m)
Max 𝐷𝑡𝑟(m) (SER=1%)
SIR (dB)

4
350
4.5

6
260
7.2

8
200
9.8

10
150
11.5

12
120
13.3

16
50
16.1

baseline methods becomes larger than 10% when 𝐷𝑡𝑟 = 250𝑚. As shown in Figure 5.9c, we can

also observe that all symbols can be decoded successfully when 𝐷𝑡𝑟 = 50𝑚. When the 𝐷𝑡𝑟 = 100𝑚,

the SER of AeroEcho reduces to 0.008, greatly lower than that of Prism and P2LoRa, which are

0.05 and 0.045, respectively. Finally, the SER of AeroEcho and two baseline methods rises to more

than 10% when 𝐷𝑡𝑟 (cid:21) 200𝑚. Based on the results, we also measure and find the maximal 𝐷𝑡𝑟 of

SER=1% with 6m, 10m, 16m 𝐷𝑡𝑟, as shown in Table 5.2. This helps the experiments for the UAV

routing scheme.

Remark: In conclusion, the performance regarding SER under different communication distances

of a single backscatter tag of AeroEcho at TVWS bands is greater than that of two baseline backscat-

ter techniques operating at ISM bands.

5.5.2 Excitation Cell

We conduct trace-driven experiments to explore the impact of maximal concurrency on ex-

citation cell radius (maximal 𝐷 𝑠𝑡) under massive concurrent collision. Different excitation cell

radii lead to different SIR (Signal-to-interference-ratio) ranges. SIR range refers to the ratio of

the strongest and weakest signals. We collect non-linear chirps in real environments and do large-

scale collision emulation with different SIR to determine maximal concurrency under different SER

thresholds.

SIR Experiments: We use fixed 𝐷𝑡𝑟 and then we move UAV to different relative distance to tag

with fixed location, making different excitation cell radius with maximal horizontal 𝐷 𝑠𝑡 from 0 to

4m, 6m, 8m, 10m, 12m or 16m. The experimental deployment is shown in Figure 5.10. Gateway

is located 100m away from tag. In the beginning, the UAV is right above the tag. Then UAV moves
right between the tag and gateway. Afterward, UAV moves circularly with step 𝜋

2 to three other

locations. We measure the SIR(maximal SNR variation) in different configurations. The results

113

Figure 5.10 The deployment of different excitation cell traces collection (drone’s view).

(a) SERthreshold=0%

(b) SERthreshold=1%

(c) SERthreshold=10%

Figure 5.11 Impact of excitation cell radius to concurrency of AeroEcho and linear chirp system.

are shown in Table 5.2. We can observe that SIR increases with 𝐷 𝑠𝑡 increasing.

Concurrency Experiment Settings: We collect accurate signals from different locations with di-

verse channels to conduct large-scale emulation of massive collisions. We improve the link diversity

by varying SIR randomly in six ranges as listed in Table 5.2. We also adopt a random time delay

among symbols ranging in [0,1] symbol time. We add multiple symbols from 2 to 100 and then

decode each symbol using a non-linear down-chirp template with sliding windows. Two kinds of

non-linear quadratic chirp and linear chirp are used. The math abstract formula of two non-linear
chirps are non1- 𝑓 „𝑡” = (cid:0) 𝐵𝑊
2

22𝑆𝐹 𝑡2 and non2- 𝑓 „𝑡” = (cid:0) 𝐵𝑊

‚ 𝐵𝑊 2
2𝑆𝐹 (cid:0)1

22𝑆𝐹 𝑡2.

𝑡 (cid:0) 𝐵𝑊 3

‚ 𝐵𝑊 3

2

Results:

Figure 5.11 shows the concurrency capacity with different SER thresholds. In Figure 5.11, it

is evident that linear chirps suffer from severe collisions. Linear chirp-based backscatter cannot

support concurrent transmission. When the excitation cell radius is 4m, quadratic1, and quadratic2

can successfully decode all the data bits when concurrency is 40, as Figure 5.11a shows. The two

can achieve 7 and 6 tags when the excitation cell radius is 16m. Figure 5.11b illustrates the overall

114

UAVTag468101216Excitation Cell Radius(m)01020304050ConcurrencyLinearQuadratic1Quadratic2468101216Excitation Cell Radius(m)020406080ConcurrencyLinearQuadratic1Quadratic2468101216Excitation Cell Radius(m)020406080100ConcurrencyLinearQuadratic1Quadratic2Table 5.3 Throughput comparison between different LoRa backscatter techniques and combinations

Netscatter P2LoRa

Prism +
P2LoRa

AeroEcho
(1%)

AeroEcho
(10%)

Throughput
Tag data rates
Concurrency

1.95kbps 2.82kbps 9.77kbps 20.5kbps
366bps
30.5bps
70
64

30.5bps
100

30.5bps
400

33.0kbps
330bps
125

AeroEcho
(10%)
+Prism

132kbps
330bps
500

AeroEcho
(1%) +
P2LoRa

1.03Mbps
366bps
3500

AeroEcho
(1%) +
P2LoRa
(SF=10)
1.46Mbps
1220bps
1500

concurrency with SER=1% is larger than the concurrency without errors at each radius. Quadratic1

and quadratic2 support (69,71), (42 40), (10 11) concurrent transmission with radius=4m, 8m, and

16m, respectively. We can also observe that they can support more than 100 concurrency when

radius=4m and SER=10%. Even with the largest radius of 16m, they can achieve 19 concurrency.

Prism [205] can only support 7 tags with SER=1% due to the availability of only 7 chirp types.

AeroEcho is 10(cid:2) the concurrency capacity of Prism and 70(cid:2) of linear chirp-based backscatter

system.

Remark: When the excitation cell radius grows, the SIRs among different backscattered signals

will also increase, causing less concurrency. A larger radius also means faster coverage and a shorter

UAV range. We carefully select the radius to balance coverage density and UAV range.

5.5.3 Throughput and data rates

Experiments Settings: To evaluate the current transmission throughput performance of ex-

isting backscatter techniques, we conduct a series of experiments at TVWS bands. As shown in

Table 5.3, we emulate the overall throughput and single tag data rates under various concurrency

values with their theoretical maximum concurrency. Our experiments involve multiple setups. We

use P2LoRa and Netscatter for basic throughput measurement with concurrency 64 and 100 respec-

tively. We integrate Prism with P2LoRa, expanding to 100 frequency channels and incorporating

4 non-linear chirp types. We also measure the performance of AeroEcho with two different SER

thresholds (1% 10%) with 70 and 125 concurrency respectively. Moreover, AeroEcho (SER=10%),

combined with Prism, creates orthogonal logical channels by assigning 125 random offsets across 4

non-linear chirp types. We combine AeroEcho (SER=1%) with P2LoRa and implement 70 random

time offsets with a single non-linear chirp type over 50 frequency channels. These experiments

115

aim to thoroughly evaluate the performance of backscatter communication systems under different

configurations and channel conditions. Sf=12 and BW=125kHz.

Results: As Table 5.3 shown, maximum concurrency of Netscatter is 64 with 125KHz. P2LoRa

takes 100 frequency channels to achieve only 2.82kbps. Prism + P2LoRa extends the orthogonal

channels by non-linear chirp type but is still restricted by the frequency channels. The first three

methods have the same data rate for individual tags: 30.5bps. When combining AeroEcho with

multiple non-linear chirp types or orthogonal frequency channels, they can all achieve the same

366bps single tag data rate, 12(cid:2) of the conventional methods. The overall throughput of AeroEcho

(1%) + P2LoRa can be up to 1.03Mbps. When the SF=10, we emulate experiments with 50 non-

linear chirps with random offsets from 50 tags in 30 orthogonal frequency channels, and the overall

throughput can be up to 1.46Mbps, which is 5.84(cid:2) the max throughput of the previous backscatter

system (250bkps of Netscatter [169]). In addition, we also verify the single tag rate can be up to

1.46kbps when bandwidth is 500kHz and SF is 12. This indicates that the flexible non-linear CSS

modulation enables AeroEcho to encode more data for each symbol or tag, which supports higher

overall throughput than previous methods. In addition, the data communication performance of

AeroEcho can be easily extended when combined with orthogonal frequency methods or multiple

non-linear chirp types methods without extra loss.

5.5.4 Aerial Routing Scheme

In this section, we compare the energy consumption of tags between the rectangular scheme

and the annular scheme. We also compare the range efficiency of UAV between fixed radius and

adaptive radius of annular AeroEcho.

Energy Experimental Settings: According to the findings detailed in Section 5.5.1, we determined

the maximum transmission distance (𝐷𝑡𝑟) across varying radii of excitation cells. We adopted a

1% SER as the reliability benchmark for the data collection system. In the rectangular scheme, the

count of columns and rows varies as integers from 1 up to K, with K signifying the point at which

the furthest excitation cells attain their maximum 𝐷𝑡𝑟 for a given radius. Similarly, we define the

number of concentric circles as integers from 1 to N for the annular scheme, where N indicates when

116

(a) Excitation cell radius=4m

(b) Excitation cell radius=8m

(c) Excitation cell radius=12m

Figure 5.12 Comparison of backscatter tags energy consumption with different excitation cell radii.

the outermost excitation cells reach their peak 𝐷𝑡𝑟. Both schemes maintain an identical density of

tag distribution. We calculate the total coverage area using each method’s maximum 𝐷𝑡𝑟. This

enables us to simulate the overall trigger frequency of backscatter tags per square meter as the

energy consumption metric.

Results: Figure 5.12 illustrates the energy consumption differences between the annular and rect-

angular schemes. At a 4m radius, the annular scheme maintains low energy consumption at ap-

proximately 1.37, while the rectangular scheme’s energy consumption spikes to 2 at a coverage

area of 7300m2, remaining around 2.1 times/m2 for larger areas—about 1.5 times higher than the

annular scheme. At an 8m radius, the annular scheme’s energy consumption is roughly 0.21, with

the rectangular scheme’s trigger times per square meter reaching 0.316, indicating slower energy

growth. Coverage is constrained for a 12m radius due to a significantly shorter maximum 𝐷𝑡𝑟, with

the annular scheme at 0.04 times/m2 and the rectangular scheme increasing from 0.045 to 0.057.

The energy disparity between the two schemes grows as the coverage area expands.

Remark: The study reveals that while the rectangular scheme leads to more incredible energy waste

for backscatter tags than the annular scheme, it enhances UAV range efficiency. Thus, the annular

scheme is preferred for energy-sensitive backscatter tags, and the rectangular scheme is better for

optimizing the UAV range.

Coverage Rate Experiments Settings: To verify the performance of coverage reliability of the

annular coverage scheme, we simulate the coverage rate for each annulus from inside to outside and

compare it with the rectangular coverage scheme.

Results: Figure 5.13 illustrates the coverage rate of two rounds for annulus 2 to 44 from inside to

117

00.511.522.533.54Coverage(m2)10511.21.41.61.822.22.4Energy(times/m2)Rectangular)Annular02468101214Coverage(m2)1040.150.180.210.240.270.30.33Energy(times/m2)Rectangular)Annular012345Coverage(m2)1040.030.0350.040.0450.050.0550.06Energy(times/m2)Rectangular)AnnularFigure 5.13 Coverage rate at the 1𝑠𝑡 and 2𝑛𝑑 rounds.

Figure 5.14 Coverage rate at 4 rounds.

outside. We can observe that the rectangular scheme achieves full coverage. In the first round, the

coverage rate of annular mostly reaches 75% to 80%. In the second round, the most annulus can

achieve more than 97% coverage rate except the second annulus. Figure 5.14 shows the coverage

rate for the second and third annulus from 1–4 rounds. The second annulus covers 97.75% tags,

and the third annulus covers 99.84% tags in the fourth round, respectively. Both can achieve 96%

in the third round.

Range Eﬀiciency Experimental Settings: To satisfy different sensor coverage densities, we need

to select an adaptive cell radius to achieve the maximal range efficiency for UAVs. We compare

adaptive radius with fixed radius to conduct experiments with different density levels and coverage

radii. Density is 1 tag/m2. The radius of the coverage area is 350m.

Results: As Figure 5.15 illustrates, the range efficiency of the two methods has the largest difference

of 0.04m/tag when the coverage radius is 350m. With coverage increasing, the available excitation

cell radius reduces, and the difference between the two methods decreases. The range efficiency is

118

01020304050Annulus7580859095100Coverage Rate(%)Annular(first round)Annular(second round)Rectangular1234Rounds7580859095100Coverage Rate(%)Annular(second annulus)Annular(third annulus)RectangularFigure 5.15 Range efficiency of fixed radius and adaptive radius.

Figure 5.16 UAV range efficiency with radius adaption and tags density.

0.037 when the coverage radius is 50m, which means only an excitation cell with a 16m radius can

be used. This implies that AeroEcho performs better in range efficiency with more excitation cells

with different radii and further coverage area. Figure 5.16 illustrates maximal density for excitation

cells with different radii. From density levels one to six, we have 1 to 6 available excitation cell

radii, respectively. The range efficiency at density level 1 and density level 6 are 0.2185 and 0.0012,

respectively. It is obvious that range efficiency decreases from density level 1 to 6, and it decreases

more and more slowly.

Remark: Adaptive excitation cell radius can achieve better range efficiency than inflexible settings.

5.6 Related Work

Long Range Backscatter: Talla et al. [167] employ a tone signal to create a linear LoRa packet

by shifting frequency. PLoRa [166] manipulates signals with two distinct frequency shifts. Uti-

lizing special excitation signals, Netscatter [169] decodes numerous LoRa packets simultaneously

by merging CSS modulation with OOK. PACT [214] enables concurrent transmission but relies on

119

350m250m200m150m120m50mCoverage Radius0.020.040.060.080.10.120.140.16Range Efficiency(m/tag)AeroEchoAeroEcho(Fixed radius)00.050.10.150.20.250.3Range Efficiency(m/tag)123456Density Level00.511.522.53Density(tag/m2)Density   AeroEcho AeroEcho(radius=4)expensive hardware each priced between 50-290 USD. P2LoRa [168] uses the existing LoRa sig-

nals for parallel decoding but requires substantial bandwidth and frequency resources. Prism [205]

achieves concurrency with different non-linear chirps to create a limited orthogonal coding space.

Contrastingly, AeroEcho proposes the using same type of non-linear chirps and mobile excitation

to improve concurrency and scalability further with low overhead.

LoRa Collision Resolving: LoRaWAN, utilizing the ALOHA protocol [185], experiences collision

when multiple nodes transmit simultaneously. Various solutions, such as Choir [186], FTrack [187],

and CIC [188], resolve these collisions by extracting unique features from overlapping packets

in the time or frequency domain. Others such as mLoRa [189], CoLoRa [215], Pyramid [191],

NScale [192], and PCube [216] use interference cancellation, spectral peak ratios, energy peak

tracking, peak scaling factor variation, and unique phase utilization, respectively. However, they

lack a collision-resolving mechanism and suffer from the near-far problem [1]. CH-MAC [217]

uses coding and hopping, and LMAC [194] attempts to avoid collisions using CAD and CSMA

at the MAC layer, but they are energy-intensive for LoRa backscatter systems. AeroEcho employs

non-linear chirps with backoff to create new logical channels, offering a practical solution for con-

currency.

UAV Routing for Backscatter system: Yang et al. [218, 219] use the UAV as receiver and provide

a multiple access solution for time division. Han et al. [220] optimize the trajectory by detecting

the presence of parasite devices. Previous studies offered general UAV-aided backscatter solutions,

which lacked specificity for agricultural contexts to solve the high cost and scalability issue. AeroE-

cho introduces an energy-efficient backscatter system for reliable data collection in agricultural IoT

environments.

5.7 Conclusion

To conclude, we develop a TVWS long-range backscatter system AeroEcho for agricultural IoT

scenarios by using non-linear chirps to enable concurrent transmissions and a UAV as a mobile exci-

tation source. We design the excitation signals on UAVs and modulation methods on backscatter tags

to improve the throughput of concurrent transmission by setting different time offsets among con-

120

current backscatter tags. We also adopt a non-linear SFD to synchronize backscatter signals from

multiple tags at the gateway side. The routing scheme achieves the balance of energy/coverage

efficiency and UAV flight range. We implement AeroEcho with customized low-cost hardware

and software-defined radio. We evaluate its performance with real environment signals. The re-

sults show that 71 tags can transmit concurrently with and less than 1% bit error rate by using the

same non-linear chirp in the same channel, resulting in a 10(cid:2) higher transmission concurrency

than state-of-the-art. Moreover, AeroEcho improves the overall throughput of current backscatter

transmission by 5.84(cid:2) and individual tag data rate by 12(cid:2) compared to the state-of-the-art.

121

CHAPTER 6

CHIRPTRANSFORMER: VERSATILE LORA ENCODING FOR LOW-POWER
WIDE-AREA IOT

6.1

Introduction

Long Range (LoRa) technology has emerged as a highly promising solution for seamlessly con-

necting unattended Internet-of-Things (IoT) devices on a large scale [7, 22]. As of March 2023, the

global presence of LoRa networks has expanded significantly, with 181 public operators worldwide

facilitating connectivity for over 300 million IoT devices [221]. This wide-area IoT technology has

found application in a multitude of scenarios. For instance, Amazon Sidewalk [222] employs LoRa

to bridge connections with smart IoT devices located beyond the range of conventional home Wi-Fi

networks. In the agricultural sphere, Microsoft FarmBeats [223, 224] utilizes LoRa to efficiently

gather data from remote sensors deployed across vast farmlands. Additionally, LoSee [95,225,226]

leverages LoRa to track shared bicycles within urban environments.

Figure 6.1 shows the network architecture underlying LoRa systems for IoT data collection [22].

LoRa nodes are responsible for encoding sensory data and transmitting data packets to LoRa gate-

ways. These gateways serve as crucial intermediaries; upon receiving the packets, they decode the

data and subsequently relay it to network and application servers through backhaul networks. It

is important to note that LoRa nodes, equipped with low-cost commercial-off-the-shelf (COTS)

Semtech LoRa radios [134, 227], operate under constrained resources.

In stark contrast, LoRa

gateways and servers utilize potent COTS Semtech radio [112, 113], some even employing Soft-

ware Defined Radio (SDR) [152], to manage energy-intensive and computationally demanding

tasks. Furthermore, distinct LoRa applications exhibit diverse performance considerations. For

example, agricultural and urban deployments prioritize ubiquitous coverage [11, 92, 95, 136, 224],

whereas industrial dense deployments contend with potential packet collisions necessitating scal-

able throughput [171, 228, 229]; In mobile and semi-outdoor applications (e.g., transportation,

elder-care), the dynamic environmental conditions can swiftly deplete the energy of LoRa nodes,

underscoring the need for sustainable adaptation strategies [230–232]. Consequently, existing

122

Figure 6.1 An illustration of the ChirpTransformer encoding framework in LoRa systems.

works predominantly focuses on optimizing network coverage [40, 100, 155, 232–237], through-

put [189, 191, 229, 238–242], and energy efficiency [230, 243–246] at the gateway and server side,

often leaving LoRa nodes untouched due to their resource limitations.

However, the efficacy of decoding designs in these systems heavily relies on how LoRa nodes

encode data. LoRa employs Chirp Spread Spectrum (CSS) modulation, where data is encoded by

linear chirps. The standard LoRa encoder uses a configurable parameter known as the spreading

factor (SF) to determine the data rate of a chirp symbol. Notably, all chirp symbols within a LoRa

packet adhere to the same SF setting. With COTS LoRa nodes offering only six available SFs

ranging from 7 to 12 for configuring LoRa encoding, the standard LoRa encoder’s reliance solely

on packet-level SF configuration presents limitations in optimizing end-to-end performance across

diverse application scenarios concerning network coverage, throughput, and energy efficiency.

In this work, we present an innovative approach to LoRa encoder design aimed at enhancing the

versatility and overall performance of LoRa networks. Our proposed solution, ChirpTransformer,

reimagines the conventional LoRa encoding framework by introducing a set of four configurable

chirp features. These features enable a more extensive range of encoding options at LoRa nodes,

as depicted in Figure 6.1. Through the utilization of these configurable parameters, a diverse array

of encoding techniques can be devised and tailored to meet the specific coverage, throughput, and

energy efficiency requirements of diverse applications. Furthermore, our approach extends beyond

node-side enhancements; we have developed corresponding decoding methodologies for gateways

123

AppEncoderLoRa NodeGateway and ServerDecoderSF conﬁguration(7 ~ 12)Standard LoRaChirpTransformerCoverage ConﬁgurationThroughput ConﬁgurationEnergy ConﬁgurationConﬁgurable Chirp FeaturesOn-air TimeSelective Initial FrequencyChirp RepeatingSymbolHoppingApplication Speciﬁc ConﬁgurationAppAgricultureIndustrySmart CityTransportationSidewalkElder CareExisting WorksNetwork Coverage, Throughput, Energy EﬃciencyCoverage DecoderThroughput DecoderEnergy DecoderApplication DecoderSmartphonePadLaptopDesktopand servers. These decoding methods leverage the pre-configured chirp features implemented by the

encoders, aiming to optimize the end-to-end performance of the LoRa network comprehensively.

Challenge #1: Chirp Feature Design. Our objective is to construct a set of specific chirp fea-

tures that collectively create an informative feature space. However, the pursuit of achieving this

goal does not involve indiscriminately adding numerous chirp features, as this would unnecessarily

complicate the LoRa network stack. To tackle this challenge, ChirpTransformer goes beyond solely

incorporating time and frequency domain data; it strategically designs four distinct chirp features

aimed at enhancing the informativeness of the feature space. First, ChirpTransformer controls the

on-air time of a symbol, which encompasses one or more chirps. Second, ChirpTransformer selects

a collection of initial frequencies that a chirp can employ to encode data. Third, ChirpTransformer

involves a novel intra-symbol chirp pattern. Departing from the conventional approach of com-

prising a symbol with just one chirp, a ChirpTransformer symbol may consist of repeated chirps

sharing the same initial frequency offset. Fourth, ChirpTransformer incorporates an innovative

inter-symbol chirp pattern. In contrast to using identical types of chirps, ChirpTransformer em-

ploys diverse chirp patterns across different symbols within a LoRa packet.

Challenge #2: Implementation on COTS LoRa nodes. Numerous chirp features, such as non-

linearity [12, 171, 247] and interleaving [248], pose implementation challenges on COTS LoRa

nodes, necessitating costly node replacements, especially in large-scale deployments. To circum-

vent this issue, the generation of these designed chirp features must not require hardware modifi-

cation or additional energy consumption. To address this challenge, we have devised a lightweight

symbol converter that transforms the chirp features – such as on-air time, selective initial frequency,

and chirp repeating – into corresponding chirps configured for SF. Furthermore, we leverage an

existing hardware interrupt specifically designed for frequency hopping to incorporate our symbol-

hopping chirp feature without incurring extra overhead.

Challenge #3: Application-specific Encoder-Decoder Co-design. With our suite of four chirp

features, we unlock the potential for multiple encoder configurations to enhance application-specific

performance. To maximize these performance gains, our focus lies in optimizing performance

124

through an end-to-end approach, encompassing encoder configuration and decoder co-design. To

address this challenge, we have undertaken three case studies targeting network coverage, through-

put, and energy efficiency:

1) Network Coverage: Addressing weak signals is crucial [95, 100]. The standard LoRa encoder

extends on-air time to ensure adequate energy reception, countering noise interference. As the

SF increases, the required signal-to-noise-ratio (SNR) for successful decoding decreases. Chirp-

Transformer takes a different approach, enhancing noise tolerance by widening the feature distance

between adjacent symbols and leveraging a neural-enhanced decoder to maximize SNR gain (x6.4).

2) Network Throughput: Enabling concurrent transmission [171, 229, 241] is a primary concern.

While LoRa permits simultaneous transmissions with different SFs, the varying on-air times across

these SFs result in imbalanced SNR tolerance among concurrent transmitters. ChirpTransformer

utilizes intra-symbol chirp repeating to create novel orthogonal encoding configurations and em-

ploys template-based decoding to resolve collisions (x6.5).

3) Network Energy Eﬀiciency: Fine-tuning encoding configurations for optimized data rates under

various noise levels is crucial [230, 243]. The standard LoRa encoder supports only six data rates,

whereas ChirpTransformer achieves 23 configurations by adjusting the number of available initial

frequencies using a selective peak searching algorithm for decoding. (x6.6).

We have implemented end-to-end ChirpTransformer systems for each case study, utilizing COTS

LoRa nodes and a USRP N210 SDR. Using our expansive campus-scale testbed spanning 2800 m

(cid:2) 1700 m, we have conducted extensive experiments to evaluate ChirpTransformer’s performance

against both standard LoRa and a state-of-the-art benchmark. The results confirm ChirpTrans-

former’s superiority across all three case studies.

In summary, our contributions are listed as follows:

(cid:15) We propose ChirpTransformer, an innovative LoRa encoding framework incorporating four chirp

features. This framework significantly enhances encoding methods, enabling efficient adaptation

to diverse applications.

(cid:15) ChirpTransformer seamlessly integrates with COTS LoRa nodes without introducing additional

125

Figure 6.2 The illustration of (a) shifted initial frequency encoding and (b) energy peak detection
decoding.

overhead. Through three comprehensive case studies, we showcase its ability to optimize end-to-

end network performance by adopting an encoder-decoder co-design.

(cid:15) We implement ChirpTransformer on COTS LoRa nodes and conduct extensive experiments us-

ing a campus-scale testbed. The results demonstrate ChirpTransformer’s exceptional performance,

achieving a 2.38 (cid:2) increase in network coverage, a 3.14 (cid:2) improvement in network throughput, and

a 3.93 (cid:2) battery lifetime compared to the standard LoRa.

6.2 Background and Motivation

6.2.1 Standard Encoding and Decoding

Figure 6.2 illustrates the encoding and decoding method of the standard LoRa. As shown in

Figure 6.2(a), the standard LoRa encoder uses bandwidth (BW) to configure a base up-chirp (e.g.,
the orange chirp), whose frequency increases linearly from (cid:0) 𝐵𝑊
2

2 over time. Notably, the

to 𝐵𝑊

on-air time of a base up-chirp is adjustable. For example, the blue base up-chirp 2 spends more

time swiping the whole BW than the orange base up-chirp 1. LoRa uses SF configuration, ranging

from 7 to 12, to control the on-air time. Given the pre-configured 𝐵𝑊 and 𝑆𝐹, the on-air time is
2𝑆𝐹
𝐵𝑊 . Thus, the on-air time is doubled when SF increases by 1.

Given the base up-chirp, data bits are encoded by shifting the initial frequency of a base up-

chirp to 𝑓0. For example, in Figure 6.2(a), we shift the initial frequency of the blue base up-chirp

126

+𝑩𝑾𝟐−𝑩𝑾𝟐𝒇𝒕𝒇𝟎𝑭𝑭𝑻. AbsFrequency Bins(a) Encoding(b) DecodingBase Up-chirp 1 Base Up-chirp 2 Encoded Chirp Symbol NoisePeakEnergy Peak𝒇𝟎2 to 𝑓0. When the chirp frequency meets BW(cid:157)2, the rest of the chirp will be shifted down and

restarts from (cid:0)BW(cid:157)2. As a result, the blue dashed chirp forms a typical chirp symbol. Moreover,

The range of 𝑓0 is »0, 𝐵𝑊”. The number of data bits encoded by a chirp symbol is also determined

by the frequency chip, indicating the distance between two adjacent initial frequency offsets. On
COTS LoRa nodes, the frequency chip is 𝐵𝑊
2𝑆𝐹 . Thus, in »0, 𝐵𝑊”, we have 2𝑆𝐹 initial frequencies to

represent data. Given an 𝑆𝐹 configuration, a chirp symbol encodes 𝑆𝐹 data bits.

At the decoder side, the dechirp [100, 241] is the standard decoding process in LoRa. First, a

received chirp symbol is multiplied with a base down-chirp (i.e., the conjugate of a base up-chirp).

The Fast Fourier Transform (FFT) is then used to focus on the energy of the chirp symbol at a single

tone, which corresponds to the initial frequency offset of the chirp symbol, on the spectrum [238,

241]. For example, in Figure 6.2(b), the blue dashed arrow indicates the energy peak by applying

FFT on the encoded chirp symbol shown in Figure 6.2(a). Then, we find the frequency bin where the

highest spectral energy peak appears to determine the initial frequency offset 𝑓0. Once we obtain

the initial frequency offset, data bits can be decoded. On the other hand, to decode a received

chirp symbol successfully, the energy peak of the received chirp symbol should be higher than the

highest energy peak derived from noises on the spectrum. For example, in Figure 6.2(b), the blue

dashed arrow is higher than the grey arrow, indicating the noise peak, to guarantee 𝑓0 can be found

correctly.

6.2.2 SF-configured LoRa Performance

We can see that the standard LoRa encoder is purely controlled by SF configuration. We have

six SF configurations, ranging from 7 to 12, on COTS LoRa nodes [134, 227]. In a LoRa packet,

all chirp symbols employ the same SF configuration for simplicity. As such, LoRa’s performance

highly depends on SF configuration. We analyze the influence of SF configurations for SNR toler-

ance and data rate as examples.

SNR Tolerance: SNR tolerance indicates the SNR threshold, above which a chirp symbol can be

successfully decoded. SNR tolerance determines the capability of network coverage. As the on-air

time of a chirp symbol increases, more energy can be combined to form a higher energy peak on

127

(a) SNR Threshold

(b) Data Rate

Figure 6.3 The influence of SF configurations on (a) SNR tolerance and (b) data rate.

the spectrum, making it easy to detect the energy peak under strong noises. Thus, a chirp symbol

with larger SFs can achieve better SNR tolerance. We set 𝐵𝑊 as 125 kHz and empirically measure

the required SNR of the dechirp under different SFs with a synthesis dataset [100,154]. Figure 6.3a

shows the SNR threshold of different SF configurations, taking 1% symbol error rate (SER) as the

criteria of successful decoding. The minimum SNR threshold reaches -22.4 dB.

Data Rate: Data rate indicates how many data bits can be successfully transmitted per second

without collision. Given a certain SNR level, a higher data rate leads to higher energy efficiency.
Given the settings of 𝑆𝐹 and 𝐵𝑊, a chirp symbol encodes 𝑆𝐹 data bits, and its on-air time is 2𝑆𝐹
𝐵𝑊 .
Thus, the data rate is 𝑆𝐹(cid:1)𝐵𝑊

2𝑆𝐹 . A chirp symbol with a larger SF has a longer on-air time, lowering its
data rate. When 𝐵𝑊 is set as 125 kHz, the data rate under different SFs is shown in Figure 6.3b.

We can see that the data rate is reduced by 42.9%-45.7% with an SF increment. As a tradeoff,

Figure 6.3a shows that increasing SF by one can achieve a 2.6-4.2 dB gain of the SNR threshold.

6.2.3 Motivation

The purely SF-controlled LoRa encoder is simple, providing SNR tolerance as low as -22.4 dB

and six data rate options under six SNR thresholds. However, with SF-12 configuration, the LoSee

measurement study [95] has shown that although the longest communication range can reach 3.2 km

- 3.5 km, a gateway can only cover about an irregular 11 km2 - 12 km2 area in an urban environment,

128

-9.8-12.8-15.6-18.2-22.4-6.2SF7SF8SF9SF10SF11SF12-25-20-15-10-50SNR Threshold(dB)SF7SF8SF9SF10SF11SF1202000400060008000Data Rate (bits/s)Figure 6.4 The illustration of the four chirp features in ChirpTransformer.

which is far from needed to achieve ubiquitous wide-area coverage. On the other hand, in those

environments with dynamic link budgets [230, 231], only six SF options are too coarse-grained

to achieve sustainable energy efficiency. This motivates us to rethink the LoRa encoder design,

particularly on COTS LoRa nodes, for inherently supporting various LoRaWAN applications.

6.3 Design of ChirpTransformer

6.3.1 Chirp Feature Design

Figure 6.4 illustrates the four chirp features, representing the information from four domains,

that are used to define a symbol in ChirpTransformer.

First, we define a time-domain feature, on-air time, which defines the propagation time of a

symbol. A symbol contains one or multiple chirps, and the total propagation time is its on-air time.

As shown in Figure 6.4(a), Symbol 1 contains one base up-chirp, and Symbol 2 has longer on-air

time with two base up-chirps.

Second, we select a set of frequencies in the range »0, 𝐵𝑊” to define the available initial fre-

quency offsets. In Figure 6.4(b), an initial frequency offset indicates the shape of a chirp compared

to the base up-chirp. Therefore, the selective initial frequencies are frequency-domain features to

determine the shapes of those chirps used to encode data.

Third, given the on-air time of a symbol, we design an intra-symbol chirp pattern, called Chirp

Repeating, to depict the repeated identical chirps in a symbol. As shown in Figure 6.4(c), Symbol

1 consists of four repetitive base up-chirps, and Symbol 2 contains two repetitive chirps with the

129

(c) Chirp Repeating(b) Selective Initial FrequencyBW/2-BW/2BW/2-BW/2(a) On-air Time(d) Symbol HoppingSymbol 1Symbol 2same initial frequency offsets. The chirp repetition is a time-domain pattern, while the same initial

frequency offset of those chirps reflects the frequency characteristic.

Lastly, given a fixed on-air time for all symbols in a packet, we design an inter-symbol chirp

operation, called Symbol Hopping, to create a new pattern domain for encoding. The basic idea

is to hop different chirp repeating patterns among the symbols in a packet instead of using the

same configuration. As shown in Figure 6.4(d), Assume Symbol 1 and Symbol 2 are in the same

packet. Symbol 1 has two repetitive base up-chirps, but Symbol 2 revises the chirp repeating pattern

with only one base up-chirp. Based on the four chirp features, we abstract four parameters to

configure encoders versatilely: 1) on-air time (OT), 2) available initial frequency offsets (IFO), 3)

chirp repeating times (CRT) in a symbol, and 4) the number of available chirp patterns in symbol

hopping (SH). ChirpTransformer adjusts the four-configuration-tuple (OT, IFO, CRT, SH) to cope

with a specific performance demand.

6.3.2 Chirp Feature Configuration

On COTS LoRa nodes, we have six types of SF-configured chirps. Based on these chirps, we

illustrate the supported configurations of our four chirp features.

OT indicates the on-air time of a symbol. On COTS LoRa nodes, the on-air time of a chirp

is determined by six SFs from 7 to 12. Its on-air time is 2𝑆𝐹

𝐵𝑊 , where 𝐵𝑊 is the bandwidth. The

minimum on-air time is with SF-7. Similar to the way of using SF to determine on-air time, our OT
is in the range of »7, ‚1”. When OT is 𝑘, the on-air time of a symbol is 2𝑘

𝐵𝑊 . In ChirpTransformer,

the on-air time of a symbol could be extended flexibly. Since the maximum SF is 12, a symbol has

to contain multiple chirps when OT is larger than 12.

IFO indicates the available initial frequency offsets for encoding. In SF-configured encoding,

an SF-𝑘 chirp is used to encode 𝑘 data bits. In ChirpTransformer, given OT-𝑘 chirps, the value of

IFO can be 2𝑖, where 𝑖 is in the range of »1, 𝑘…. Thus, the configuration (OT-𝑘, IFO-𝑖) means that

we use OT-𝑘 chirps to encode 𝑖 data bits by shrinking the available initial frequency offsets from 2𝑘

to 2𝑖. For example, the configuration (OT-7, IFO-6) indicates using OT-7 chirps to encode 6 data

bits. Half of the initial frequency offsets in OT-7 chirps will no longer be used for encoding. When

130

Figure 6.5 Two examples to illustrate the concepts of CRT and SH. (a) parameter tuple (OT-10,
IFO-9, CRT-2); (b) parameter tuple (OT-10, IFO-1, SH-2).

OT is larger than 12, the maximum IFO can be 212 on COTS LoRa nodes. In such cases, the IFO

values are from IFO-1 to IFO-12.

CRT indicates the number of repetitive identical chirps in a symbol. We know that the on-

air time of an OT-10 symbol equals eight SF-7 chirps, four SF-8 chirps, two SF-9 chirps, and

one SF-10 chirp. Similarly, when the on-air time of a symbol is OT-𝑘, we use CRT- 𝑗 repetitive

SF-„𝑘 (cid:0) 𝑗” chirps with the same initial frequency to fill the symbol, where 𝑗 is in the range of

»maxf𝑘 (cid:0) 12, 0g, 𝑘 (cid:0) 7…. For example, as shown in Figure 6.5(a), (OT-10, IFO-9, CRT-1) means

the on-air time of each symbol is OT-10. Each symbol consists of two (i.e., 21) identical SF-9 (i.e.,

10 (cid:0) 1) chirps, which have 29 available initial frequency offsets for encoding.

SH represents the available chirp repeating patterns of symbol hopping. SH has three values:

0, 2, and 4. If SH is 0, the symbols in a packet follow the same pattern without symbol hopping.

The data bits encoded by a symbol purely rely on the IFO settings. If SH is 2 or 4, we have 2 or 4

different patterns. We can use the different patterns to encode 1 data bit or 2 data bits. When SH

is not 0, CRT will be invalid. Given the on-air time of a symbol OT-𝑘 (𝑘 (cid:20) 12) and the available

chirp patterns SH-𝑡, the 𝑡 chirp repeating patterns of a symbol can be represented by one SF-𝑘 base

up-chirp, or two SF-„𝑘 (cid:0) 1” repetitive base up-chirps, ..., or 2𝑡 SF-„𝑘 (cid:0) 𝑡” repetitive base up-chirps.

We must keep 𝑘 (cid:0) 𝑡 (cid:21) 7. And when 𝑘 is larger than 12, the SH-𝑡 chirp repeating patterns include

2𝑘(cid:0)12 SF-12 base up-chirps, 2𝑘(cid:0)11 SF-12 base up-chirps, ..., or 2𝑘(cid:0)13‚𝑡 SF-„13 (cid:0) 𝑡” base up-chirps.

For example, as shown in Figure 6.5(b), (OT-10, IFO-0, SH-2) means the on-air time of a symbol

is OT-10. IFO-0 indicates we do not use initial frequency offsets to encode data. In this case, with

SH-2, we have two symbols. One is two repetitive SF-9 base up-chirps, indicating bit ‘0’. The other

131

Symbol 1: OT-10(a) (OT-10, IFO-9, CRT-1)Symbol 2: OT-10SF-9SF-9SF-9SF-9IFO-9CRT-2CRT-2(b) (OT-10, IFO-0, SH-2)Symbol 1: OT-10Symbol 2: OT-10SF-9SF-9SF-10IFO-0SH-2: ‘0’SH-2: ‘1’Figure 6.6 Timeline of the symbol hopping feature, leveraging a hardware interrupt.

is one SF-10 base up-chirp, indicating bit ‘1’.

6.3.3 COTS Feature Implementation

A ChirpTransformer packet consists of four parts: preamble, start frame delimiter (SFD), header,

and payload. The format of the preamble, SFD, and header are the same as the standard LoRa. We

put the values of (OT, IFO, CRT, SH) in the header. The symbols in the payload are encoded by

the encoding method determined by (OT, IFO, CRT, SH). By default, COTS LoRa nodes support

SF-configured packets in which all chirp symbols follow the same SF configuration. However, if

symbol hopping is enabled, a ChirpTransformer packet consists of symbols with different SF chirps.

Thus, without symbol hopping, we design a symbol converter to translate the payload of a Chirp-

Transformer packet to SF-configured chirp symbols. On the other hand, we leverage a hardware

interrupt to implement symbol hopping.

Given OT-𝑘, IFO-𝑖, and CRT- 𝑗 configurations, the symbol converter includes three steps. In the

first step, it determines the corresponding SF configuration. If only one chirp exists in a symbol, the

SF configuration corresponds to OT-𝑘. Otherwise, the symbol contains a chirp repeating pattern

CRT-𝑡. Then, the SF configuration is 𝑘 (cid:0) 𝑡 corresponding to the SF of each repetitive chirp. In step

two, we calculate the number of SF chirps to form a symbol. If CRT is not applied, only one SF

chirp exists. For CRT- 𝑗, the number corresponds to 2 𝑗 . In the last step, we set the initial frequency

offset for each SF chirp. Since IFO-𝑖 must be a subset of IFO-𝑆𝐹, we directly assign the initial

frequency offset of each chirp in a ChirpTransformer symbol to the corresponding chirp generated

132

RadioFreqHoppingPeriodInterrupt SignalPreambleHeaderPayloadSFDEncoding Bits001012104118001118012104001012104...INVALIDINVALIDSet configurations for the next hoptFigure 6.7 Current profile for a ChirpTransformer transmission with symbol hopping.

in step two. In this way, a ChirpTransformer symbol can be translated to a series of SF-configured

chirp symbols.

We implement the symbol hopping feature based on the frequency hopping capability of COTS

LoRa nodes [227]. The LoRa standard requires LoRa nodes to support frequency hopping, enabling

long-duration packet transmission without violating the maximum permissible channel dwell time.

The key principle behind the frequency hopping scheme is hardware interrupt, named ChangeChan-

nelFhss, that enables LoRa nodes to select and switch to a new channel during packet transmission.

After a predetermined hopping period, the transmitter and receiver change to the next channel in

a predefined list of hopping frequencies to continue transmission and reception of the next portion

of the packet. Our key observation is that a LoRa node can modify not just the channel but all

configurations every time it triggers the ChangeChannelFhss interrupt. Thus, we can implement

symbol hopping on a COTS LoRa node without adding extra hardware by making the node peri-

odically trigger the ChangeChannelFhss interrupt and change its SF configuration during packet

transmission. Moreover, we keep the transmission on a single channel by setting the list of hopping

frequencies as identical frequencies.

For example, we use (OT, IFO-0, SH-4) to encode 2-bit data with four different chirp repeating

patterns. Figure 6.6 shows the timeline of the encoder for transmitting a packet. The transmission

starts with the preamble, SFD, and header. At the beginning of each transmission, an interrupt

signal ChangeChannelFhss is generated, where the interrupt handler programs the frequency, SF,

and hopping period for the first hop of the payload. The interrupt signal is cleared after all config-

133

Time (s)Current (mA)0501001502000.01.50.51.0MCU & Radio Sleep1.7 μAMCU + Tx~140 mAMCU only<5 mAFluctuation by FHSS interrupturations have been settled. Then, during payload transmission, the transmitter periodically triggers

the ChangeChannelFhss interrupt to modify the SF configuration for ChirpTransformer data mod-

ulation. We use our symbol converter to determine the SF for each chirp repeating pattern. The

time that each hop of transmission will dwell is determined by FreqHoppingPeriod, which is an

integer multiple of symbol periods. As illustrated in data modulation, the periods of all symbols

should be identical to the OT configuration. Thus, we determine the FreqHoppingPeriod based

on the SF configuration of chirps in the corresponding hop, i.e., 𝐹𝑟𝑒𝑞𝐻𝑜 𝑝 𝑝𝑖𝑛𝑔𝑃𝑒𝑟𝑖𝑜𝑑 = 2𝑂𝑇(cid:0)𝑆𝐹.

The new configurations are programmed within the current hopping period to ensure it has been set

when the next hop begins. The interrupt computation is much shorter than the symbol on-air time,

causing no extra latency.

6.3.4 Energy Profiling

To demonstrate the energy overhead brought by the hardware interrupt processing, we estimate

the energy profile of a ChirpTransformer transmission with symbol hopping using the Monsoon

HV Power Monitor [249]. The node is powered by 3.6 V. Figure 6.7 shows the current profile of

the transmission, including the instant current for all radio access phases. The LoRa node stays in

sleep mode when it does not transmit data. The radio transmission consumes the highest amount of

energy by a large margin. The power consumption of the MCU is much less than that of the radio

circuit. Thus, the ChangeChannelFhss interrupt processed by the MCU only introduces a small

current fluctuation during the packet transmission, which has a neglectable impact on the whole

energy consumption of the LoRa node.

6.4 Case I: Network Coverage

In the first case study, we focus on reliable weak signal decoding to extend LoRaWAN cover-

age [100]. We are seeking a design to tolerate the SNR lower than the -22.4 dB SNR threshold

of the standard LoRa (x 6.2.2). With ChirpTransformer, the key idea is to maximize the feature

distance among different symbols during encoding beyond SF-12 with SH configurations.

134

Figure 6.8 The illustration of symbol hopping encoding, using four chirp repeating patterns (e.g.,
SF-9 – SF-12) to represent 2-bit data.

6.4.1 SH Encoder-Decoder Co-design

Encoding Method: Our design utilizes the chirp pattern domain created by symbol hopping

and prohibits using initial frequency offsets (i.e., IFO-0) for data encoding. We use four chirp pat-

terns to represent each 2-bit data (i.e., SH-4). To obtain four chirp repeating patterns, the minimum

OT configuration is OT-10 so that we have the SF-7, SF-8, SF-9, and SF-10 chirps to construct

four different patterns. When OT increases, the SNR tolerance will be enhanced while the data rate

reduces. Figure 6.8 shows an example of (OT-12, IFO-0, SH-4), which uses four chirp repeating

patterns of SF-9 – SF-12 to define 4 symbols. We use SH-[9-12], SH-[8-11], and SH-[7-10] to

represent (OT-12, IFO-0, SH-4), (OT-11, IFO-0, SH-4), and (OT-10, IFO-0, SH-4), separately.

Neural-enhanced Decoding Method: Our decoding problem is to extract a symbol’s chirp patterns

(i.e., 4-class classification). Inspired by the neural-enhanced decoder in NELoRa [100], Chirp-

Transformer also aims to decode the symbols in a neural-enhanced manner to maximize the SNR

gain from both the encoder and decoder side. Like NELoRa, our neural-enhanced decoder converts

symbols into time-frequency spectrograms as input feature maps. We simplify NELoRa’s archi-

tecture by using a lightweight first Conv2d module with fewer filters and replacing the LSTM layer

with a bidirectional GRU layer, which is more computation-efficient for temporal feature extraction.

Figure 6.9 shows the concise network structure, which consists of seven modules in total. Specifi-

cally, the first four modules aim to generate a filter mask to be multiplied with the input spectrogram.

Then, we feed the masked spectrogram into a three-module classifier for chirp pattern recognition.

135

Figure 6.9 The structure of the neural-enhanced decoder.

Figure 6.10 The illustration of the IFO-2 based LoRa encoder, using four initial frequency offsets
(e.g., 𝑓0, 𝑓1, 𝑓2, 𝑓3) to represent 2-bit data.

6.4.2

IFO-2 based Encoder-Decoder Co-design

Encoding Method: With the similar design principle of our symbol hopping encoder, we can

set the available initial frequency offsets as 4 (i.e., IFO-2) to encode 2-bit data. Specifically, as

shown in Figure 6.10, we select four SF-12 chirps with deterministic initial frequency offsets (e.g.,
4 , 0 and 𝐵𝑊
𝑓0, 𝑓1, 𝑓2, 𝑓3) to encode 2-bit data. We set the initial frequency offsets as (cid:0) 𝐵𝑊
to distinguish these symbols as much as possible in the feature space. To keep a similar data rate

2 , (cid:0) 𝐵𝑊

4

to the symbol hopping encoding, IFO based encoding has three configurations, namely IFO-2-10,

IFO-2-11, and IFO-2-12, which use SF-10, SF-11, and SF-12 chirps to form symbols matching

SH-[7-10], SH-[8-11], and SH-[9-12] symbols.

Neural-enhanced Decoding Method: The decoding problem becomes classifying the initial fre-

quency offsets, which is still a four-class classification. We adopt the same deep neural network

(DNN) structure in Figure 6.9 to classify IFO-2 based symbols for decoding.

136

Feature Map2 ×64×129×4  Chirp PatternConfigurationBidirectional GRUDense + ReLUDense + SigmoidConv2d16 @ (3,3)Dense + ReLU+ DropoutDense + SoftMaxConv2d8 @ (1,7)Conv2d8 @ (7,1)Conv2d8 @ (5,5)Conv2d8 @ (5,5)Conv2d8 @ (5,5)Conv2d8 @ (5,5)Conv2d8 @ (1,1)-BW/2+BW/2TimeFrequencySymbol-1Symbol-2Symbol-3Symbol-4Configuration: Spreading Factor = 12𝑓!𝑓"𝑓#𝑓$6.4.3 Weak Signal Packet Detection

After a LoRa node transmits a packet through the air, the packet should be reliably detected

at the decoder side. Then, its payload is divided into multiple symbols and fed into the DNN-

based decoder for chirp pattern recognition. Since the preamble part of a ChirpTransformer packet

remains the same as the standard LoRa encoder, we need a reliable way to detect the LoRa preamble

under ultra-low SNR. Given a period of received signals, we identify whether they contain a LoRa

preamble, which consists of a series of base up-chirps, as an indicator of whether a packet is coming.

We divide the received signals into 𝑁 symbol-length signal segments, where 𝑁 is the number of

base up-chirp symbols in a LoRa preamble. Then, we combine the 𝑁 segments as a superposed

signal segment. All the base up-chirp symbols are constructively superposed if a preamble is in

the received signals [100]. Considering the random initial phase of each base up-chirp symbol due

to carrier frequency offset (CFO) and sampling frequency offsets (SFO), along with the quadratic

distribution of the phase shift related to frequency bias and chirp symbol index, we heuristically

determine the optimal phase compensation [100, 238]. After phase compensation, we coherently

sum up the 𝑁 symbol-length signal segments of the received signals. Then, we use a standard base

up-chirp symbol to calculate its cross-correlation with the superposed signal segment. We treat

the detection of a significant correlation peak as a sign of a successful preamble detection. The

threshold for peak detection is based on channel estimation, which is six standard deviations of

the mean noise correlation. The index of the correlation peak also indicates the boundary of the

received LoRa preamble symbols. Therefore, we can align the timing of the demodulation window

and extract the aligned symbols from the payload part using the detected preamble.

6.4.4

Implementation

We have implemented ChirpTransformer on COTS LoRa nodes and use an SDR as a gateway,

shown at the bottom of Figure 6.11. Specifically, the USRP N210 SDR platform captures over-

the-air LoRa signals by operating on a UBX daughter board with a sampling rate of 1 MS/s for

both ChirpTransformer and standard LoRa, which is a widely used sampling rate setting [40, 100,

171, 238]. Since the maximum 𝐵𝑊 is 500 kHz in LoRa, COTS LoRa gateways should support a

137

Figure 6.11 USRP N210 based gateway and COTS LoRa nodes deployed in a campus environment.

sampling rate of at least 1 MS/s according to the Nyquist–Shannon sampling theorem. The captured

signal samples are then delivered to a back-end host, preprocessed, and demodulated by the decoder

algorithms. The COTS SX1278 [227] based LoRa nodes transmit ChirpTransformer packets with

random payloads. They are deployed in a campus environment, including both indoor and outdoor

scenarios, as shown at the top of Figure 6.11. ChirpTransformer implements the DNN model on

a Raspberry PI 4 [250] with an average inference time of 0.26 s over 100 runs and a memory

requirement of 17.19 MB. In comparison, the compressed NELoRa model under the SF-10 setting

reaches 152.9 MB in memory and 0.97 s for inference. This makes NELoRa 3.7(cid:2) slower and 8.9(cid:2)

larger in memory than ChirpTransformer. This can further help reduce the overhead of the DNN

model deployment at the gateway.

6.4.5 Baseline Methods and Metrics

Besides the standard LoRa, which uses SF-12 chirps for encoding and the dechirp for decoding,

we choose Ostinato [251], a state-of-the-art encoder-decoder co-design for weak signal decoding

beyond SF-12, as one of our baselines.

Ostinato: repetitive SF-12 chirps based encoder + chirp coherent-combining based decoder.

Ostinato [251] can be regarded as a special case of ChirpTransformer’s chirp repeating encoding.

Specifically, Ostinato uses repeated SF-12 chirps, having the identical initial frequency offset, to

encode the same data as a single SF-12 chirp does. Then, the decoder coherently combines the

138

multiple SF-12 chirps to obtain a more obvious energy peak during dechirp than any individual SF-

12 chirp. We implement the encoder by using our chirp repeating feature to generate the repeated

SF-12 chirps and adopt the phase calibration method to fine-tune the chirp coherent combining for

reliable decoding.

We use two metrics to indicate the performance of weak signal decoding.

(cid:15) Symbol Error Rate (SER) is widely used to demonstrate the channel noise resilience of a physical

layer design given different SNR [100, 238, 239]. A low SER is desirable.

(cid:15) SNR Threshold is the lowest SNR that a physical layer design can keep the SER under a prede-

termined value. By default, we set the predetermined SER as 1%.

6.4.6 Communication Reliability Evaluation

Setup: We evaluate the SNR threshold of ChirpTransformer under SH-[7-10], SH-[8-11], SH-[9-

12], IFO-2-10, IFO-2-11, and IFO-2-12 configurations. We further evaluate the SNR threshold of

Ostinato with three comparable configurations, using single, two, and four repetitive SF-12 chirps

to compose a symbol, indicated as Ostinato-1, Ostinato-2, and Ostinato-4. In the experiments, we

collect high-SNR symbols on our campus testbed. Then, we inject random noises to generate weak

signals with arbitrary SNR levels using the same method of NELoRa [100]. We use the synthesis

symbols to calculate the SNR threshold, indicating communication reliability. We generate the

same number of synthesis symbols to train the DNN decoder for our symbol hopping and IFO-2-

based encoders.

Results: The results are shown in Figure 6.12.

General SNR Gain: The SNR thresholds of symbol hopping encoder are -24.5 dB, -27.4 dB and

-28.8 dB using SH-[7-10], SH-[8-11], and SH-[9-12] encoding settings, separately. Compared with

-22.4 dB SNR threshold of the standard LoRa under SF-12, we achieve the maximum 6.4 dB SNR

gain, which is much larger than the approximate 2 dB SNR gain of NELoRa [100]. This indicates

that ChirpTransformer significantly improves the communication reliability of LoRa. LoSee [95],

an urban LoRa measurement study, introduces a link model to predict the packet delivery ratio

(PDR) within a square coverage area based on link SNR. The coverage area is defined where the

139

Figure 6.12 The comparison of SNR threshold between ChirpTransformer and Ostinato.

overall PDR value exceeds 70%. With the link model of LoSee and 6.4 dB SNR gain, ChirpTrans-

former can achieve approximately 2.38(cid:2) coverage area compared with the standard LoRa in the

urban environment.

SNR Gain from symbol hopping: the IFO-2 based encoder adopts the same DNN-based decoder

with the symbol hopping encoder but encodes data with the same type of chirps. We can see that the

SNR thresholds of the IFO-2 based encoder are 1.2 dB, 1.0 dB, and 0.9 dB higher than the symbol

hopping encoder with the identical on-air time settings of OT-10, OT-11, and OT-12. On aver-

age, the symbol hopping encoder can tolerate 1.03 dB lower SNR than the IFO-2 based encoder.

This verifies our symbol hopping feature provides a larger feature space than IFO-2 feature, which

makes the different symbols can be easily distinguished with a DNN model. The SNR thresholds

of SH-[7,10], SH-[8,11], and SH-[9-12] are 2.1 dB, 2.4 dB, and 0.6 dB lower than Ostinato-1,

Ostinato-2, and Ostinato-4, separately. On average, ChirpTransformer achieves 1.7 dB SNR gain

compared to Ostinato. This indicates our symbol hopping feature outperforms the chirp repeating

pattern in Ostinato with complicated phase calibration for chirp coherent-combining.

6.4.7 Campus-scale Evaluation

Setup: SH is adopted by those LoRa nodes experiencing weak links to improve network coverage.

The SH encoder extends its range and reduces the SNR threshold by lowering the data rate and

neural-enhanced decoder. We conduct campus-scale experiments to verify the gain of SH encoder-

140

-22.4-23.3-24.5-25.0-26.4-27.4-27.9-28.2-28.8Ostinato-1IFO-2-10SH-[7-10]Ostinato-2IFO-2-11SH-[8-11]IFO-2-12Ostinato-4SH-[9-12]-32-30-28-26-24-22-20SNR Threshold(dB)OstinatoIFO-2 EncoderSymbol Hopping EncoderFigure 6.13 The campus-scale testbed with 20 LoRa nodes at 20 indoor and outdoor locations.

decoder co-design with extremely low SNR conditions. Figure 6.13 illustrates the deployment of

our outdoor testbed at a campus (2800𝑚(cid:2)1700𝑚). We randomly deploy 20 COTS LoRa nodes at 20

different NLOS positions covering indoor and outdoor scenarios. For indoor nodes, concrete walls

are the main obstacles. Buildings and trees are the obstacles in the outdoor environment. For each

LoRa node at a position, we collect tens of packets, and each contains 40 payload symbols. We re-

peat the data collection with three encoding methods of the standard LoRa, ChirpTransformer using

SH-[9,12] configuration, and Ostinato using Ostinato-4 configuration, respectively. We first eval-

uate the preamble detection accuracy of our design compared to the standard LoRa. In addition,

for those detected packets, we compare the decoding SER between ChirpTransformer and Osti-

nato. For our SH-[9,12], we apply the DNN-based decoder pre-trained with our synthesis dataset

collected in x 6.4.6 on the weak signal packets.

Results for packet detection: To understand whether a ChirpTransformer packet can be success-

fully detected at extremely low SNR conditions, we evaluate ChirpTransformer’s performance on

packet detection for all 20 positions. Figure 6.14a shows the comparison of the packet loss rate

between ChirpTransformer and the standard LoRa. Figure 6.14b shows the CDFs of the packet

loss rate for ChirpTransformer and the standard LoRa. We can see that the standard LoRa suffers

a high packet loss rate at all positions, where more than 90% of packets are undetected at 10 out

of 20 positions. In comparison, ChirpTransformer achieves a lower than 6% packet loss rate at all

positions. This is because the standard LoRa detects LoRa preambles by searching for continu-

141

Outdoor ClientGatewayOutdoor ClientGatewayIndoor ClientOutdoor ClientGatewayIndoor Client124365100m188997788111112121313101014141919202017171111881010202016161515121211810201615121515161618182800 m(a) Cross-position Performance

(b) CDF Performance

Figure 6.14 The packet detection accuracy on our campus-scale testbed with extremely low SNR.

ous identical frequency-domain energy peaks by applying the dechirp to the continuously received

signals. Thus, it requires the energy peak of each chirp symbol in the preamble to be detectable.

However, the real SNR at most of the positions is much lower than the SF-12 SNR threshold (e.g.,

-22.4 dB), leading to undetectable energy peaks. After mitigating the CFO and SFO in a preamble,

ChirpTransformer concentrates the energy of all preamble chirp symbols to detect a LoRa packet.

The results verify that ChirpTransformer can reliably detect any LoRa packets at the extremely low

SNR.

Results for decoding SER: We utilize our preamble detection design to detect Ostinato packets.

Then, we compare ChirpTransformer with Ostinato by computing the decoding SER for the detected

symbols at all 20 positions. As shown in Figure 6.15b, we can see the median SER of Ostinato

is about 79.25%.

In comparison, ChirpTransformer can achieve a much lower median SER of

40.54%, indicating the boundary of the communication range will be enlarged greatly. The reason

is that Ostinato suffers from severe noises in the wild environment, leading to the signals not being

coherently combined. At the same time, our neural-enhanced decoder can tolerate it by exploring

multi-dimensional features. The specific SER across 20 positions is shown in Figure 6.15a. Our

SER is lower than Ostinato at all 20 positions. In addition, we observe that ChirpTransformer has

a higher SER in some positions, such as positions 4, 7, 8, 14, and 15, than others. The reason is

that some coexisting wireless interference also brings new noise patterns that the pre-trained DNN

142

5101520Position00.51Packet Loss RateStandard LoRaChirpTransformer00 0.5 1Packet Loss Rate0.51CDFChirpTransformerStandard LoRa(a) Cross-position Performance

(b) CDF Performance

Figure 6.15 The comparison of communication reliability on our campus-scale testbed.

model does not see, degrading the SER. We can involve an online fine-tuning process to deal with

those new noise patterns [100].

6.5 Case II: Network Throughput

In our second case study, we target to enable efficient concurrent transmission, allowing mul-

tiple LoRa nodes to transmit their packet simultaneously to enhance the network throughput and

scalability [171, 229, 241]. Specifically, in standard LoRa, concurrent transmission is infeasible for

multiple packets with the same SF configuration [171]. The six SFs create six quasi-orthogonal

logic channels to enable concurrent transmission [242]. However, since the SNR threshold in-

creases when SF decreases (x 6.2.2), we cannot assign arbitrary SF to a LoRa node without notic-

ing its SNR requirement, thus degrading the efficiency of concurrent efficiency. With ChirpTrans-

former, given the same SNR tolerance, we aim to design multiple orthogonal logic channels to

achieve efficient concurrent transmission. Our key idea is to use different repeating chirp patterns

with CRT configurations to create orthogonal logic channels.

6.5.1 CRT Encoder-Decoder Co-design

Encoding Method: For an orthogonal logic channel, the encoding method can be indicated

as (OT-𝑘, IFO-𝑖, CRT-„𝑘 (cid:0) 𝑖”, SH-0). OT-𝑘 is the pre-configured on-air time. IFO-𝑖 indicates

that we use SF-𝑖 chirps to construct the chirp repeating pattern and encode 𝑖 data bits. The chirp

143

5101520Position00.51Symbol Error RateChirpTransformer Ostinato00.51SER00.51CDFChirpTransformer Ostinato(a) Chirp Repeating

(b) Template Downchirp

Figure 6.16 Six orthogonal logic channels with OT-12.

repeating times are 2𝑘(cid:0)𝑖. 𝑖 is in the range of »7, maxf𝑘, 12g…, indicating 𝑘 must be larger than 𝑖 to

create a diverse chirp repeating pattern, and 𝑖 is not less than 7, the minimum SF chirp we have.

Therefore, given the pre-configured on-air time OT-𝑘, we have 𝑘 (cid:0) 7 orthogonal logic channels if

𝑘 (cid:20) 12. Otherwise, we have six orthogonal logic channels using SF-7 to SF-12 chirp repeating

patterns. The larger the 𝑘 is, the more orthogonal logic channels we can have. Figure 6.16a shows

an example of the on-air time OT-12. We have six orthogonal logic channels with 32 SF-7 chirp

repeating (OT-12, IFO-7, CRT-5, SH-0), 16 SF-8 chirp repeating (OT-12, IFO-8, CRT-4, SH-0), 8

SF-9 chirp repeating (OT-12, IFO-9, CRT-3, SH-0), 4 SF-10 chirp repeating (OT-12, IFO-10, CRT-

2, SH-0), 2 SF-11 chirp repeating (OT-12, IFO-11, CRT-1, SH-0), and one SF-12 chirp (OT-12,

IFO-12, CRT-0, SH-0).

Template Down-chirp based Decoding: Given the (OT-𝑘, IFO-𝑖, CRT-„𝑘 (cid:0) 𝑖”, SH-0), we use

a template down-chirp to find the energy peak in the FFT spectrum. The template down-chirp

consists of 2𝑘(cid:0)𝑖 SF-𝑖 base down-chirps. Figure 6.16b shows the template down-chirps for our six

logic channels with OT-12. The basic observation is illustrated in Figure 6.17. Given an (OT-

12, IFO-11, CRT-1, SH-0) symbol with two SF-11 base up-chirps, we multiply two matched base

down-chirps (called template) and coherently combine the derived signals in the two SF-11 windows

144

32X SF716X SF88X SF94X SF102 X SF111 X SF12Figure 6.17 After multiplying the matched template down-chirp, a superposed energy peak appears.

together. After applying FFT, a superposed energy peak appears at bin 0, which is equivalent to

the energy peak of an SF-12 base up-chirp due to the same on-air, thus supporting the same SNR

tolerance ability. When using a mismatched template like an SF-12 base down-chirp, the energy is

dispersed at all bins.

With this observation, the decoder first initializes maxf𝑘 (cid:0) 7, 6g down-chirp templates with

different SF configurations. Then, for each template, due to the CFO and SFO among the repetitive

chirps in the symbol, we compensate various phase offsets calculated by the CFO and SFO esti-

mated during packet detection for alleviating the random initial phase problem [100], then apply

the template to obtain the corresponding spectrum with FFT. With the phase compensation, we can

make the obtained energy peak at frequency bin 𝑓0 as accurate as possible. Suppose the derived

energy peak with a template is larger than the current maximum energy peak. In that case, we

update the expected initial frequency offset 𝑓0 and IFO configuration of the symbol and the latest

maximum energy peak. When traversing all templates, the maximum energy peak is selected, and

the 𝑓0 and IFO configuration are converted to data bits accordingly.

6.5.2

Implementation and Baseline

We implement the encoder and decoder on our campus testbed. Besides the standard LoRa,

we select CurvingLoRa [171] as the state-of-the-art baseline from the perspective of LoRa encoder

design. CurvingLoRa utilizes the orthogonal coding space created by different non-linear chirps

to enable concurrent LoRa transmissions. The non-linear chirps have similar SNR tolerance com-

pared to the linear chirp with the same on-air time. Moreover, besides SER, we use another metric

145

+𝐵𝑊2−𝐵𝑊2𝑓𝑡𝐴𝑚𝑝𝑙𝑖𝑡𝑢𝑑𝑒02!!−1𝐹𝐹𝑇𝑏𝑖𝑛Match            MismatchTwo SF-11 base up-chirps: Superposed energy peak at bin 0(a) SER

(b) Throughput

Figure 6.18 Collision performance of ChirpTransformer, CurvingLoRa and LoRa.

Throughput, which indicates the amount of data transmitted over a network in a unit of time.

6.5.3 Concurrent Transmission Evaluation

Setup: We collect high-SNR symbols from our campus testbed. We conduct trace-driven experi-

ments to evaluate the SER and network throughput of ChirpTransformer, CurvingLoRa [171], and

the standard LoRa during concurrent transmission. As the same operation in CurvingLoRa [171]

to emulate signal collisions, we collect individual symbols in real environments separately and then

add the symbols together to generate the overlapping patterns. To simulate diverse temporal patterns

of symbol collisions, we randomly assign a symbol offset from [0,1] times the symbol on-air time.

To evaluate the performance under near-far issues, we add a signal-to-interference ratio (SIR) rang-

ing from -20 to 0 dB to concurrently transmitted symbols. The frequency bandwidth is 125 kHz.

For the standard LoRa, we use SF-12 chirps for all concurrent transmitters. For CurvingLoRa, we

adopt five types of chirps: quadratic1, quartic1, quadratic1, quartic1, and linear with on-air time

OT-12. Since we assign concurrent transmitters to different orthogonal logical channels, the five

logic channels only support five concurrencies at most. For ChirpTransformer, we use the six or-

thogonal chirp repeating logic channels (OT-12, IFO-12, CRT-0, SH-0), (OT-12, IFO-11, CRT-1,

SH-0), (OT-12, IFO-10, CRT-2, SH-0), (OT-12, IFO-9, CRT-3, SH-0), (OT-12, IFO-8, CRT-4, SH-

0), and (OT-12, IFO-7, CRT-5, SH-0). To calculate the average network throughput and SER. We

repeat experiments for each concurrency configuration 2,000 times.

146

23456Concurrency10-410-310-210-1100Symbol Error RateLoRaChirpTransformerCurvingLoRa23456Concurrency050010001500Throughput (Bit/s)LoRaChirpTransformerCurvingLoRaResults for Symbol Error Rate Performance: The SER for different concurrency values from 2

to 6 is shown in Figure 6.18a. 10(cid:0)4 represents no error symbols. When concurrency is 2, LoRa

fails to decode half of the symbols correctly. In contrast, the SER for CurvingLoRa is less than

10%, and there is no error symbol for ChirpTransformer. As concurrency increases, the SER for

CurvingLoRa and ChirpTransformer also increases. However, CurvingLoRa exhibits significantly

higher symbol decoding errors compared to ChirpTransformer. Despite a SIR range of [-20,0] dB

and a concurrency of 6, ChirpTransformer can still maintain a SER less than 0.2%.

Results for Network Throughput Performance: As shown in Figure 6.18b, LoRa consistently

has the lowest throughput due to its high SER, unaffected by concurrency values ranging from 2 to

6. In contrast, ChirpTransformer consistently outperforms CurvingLoRa in throughput due to our

lower SER. Specifically, at a concurrency of 5, the throughput of ChirpTransformer is 3.33(cid:2) that

of LoRa and 1.07(cid:2) that of CurvingLoRa, demonstrating ChirpTransformer’s superior performance

in concurrent transmissions.

6.6 Case III: Network Lifetime

In the standard LoRa, we configure SF to balance between SNR tolerance and data rate. After

a LoRa node is deployed or its surrounding environment changes, a LoRa gateway will check the

node’s SNR and choose an SF configuration that can reach the highest data rate to enhance the

energy efficiency while its SNR threshold is lower than the observed SNR to keep reliable packet

delivery. Only six SFs are too coarse-grained to optimize network energy efficiency in complex

environments [230, 231]. As such, we aim to develop fine-grained data rate adaptation with IFO

configurations to achieve energy efficiency.

6.6.1

IFO based Encoder-Decoder Co-Design

Encoding Method: We adjust the available initial frequency offsets to obtain more encoding con-

figurations. The encoding method can be indicated as (OT-𝑘, IFO-𝑖, CRT-0, SH-0). For energy

efficiency, the on-air time is not larger than that of an SF-12 chirp. Thus, we have 𝑘 (cid:20) 12 and
𝑖 (cid:20) 𝑘. Given the on-air time OT-𝑘, we encode 𝑖 bits of data. The data rate is 𝑖(cid:2)𝐵𝑊
2𝑘

. To encode 𝑖

bits data, we only need 2𝑖 initial frequency offsets. Given the total 2𝑘 initial frequency offsets, we

147

uniformly remove unused offsets and select the rest as the available initial frequency offsets for en-
coding. For example, for OT-12, the range of initial frequency offsets is »0, 212(cid:0)1
212
𝐵𝑊, ..., 212(cid:0)4
212

we uniformly select a quarter of these offsets as f0, 4
212
from 7 to 11, 𝑖 is in the range of [d 𝑘‚1
2
when 𝑘 is 7, the available 𝑖 is 7, 6, and 5. If 𝑖 is 4, the data rate is 4𝐵𝑊

e, 𝑘]. When 𝑘 is 12, 𝑖 is in the range of [6, 12]. For example,
27 which is the same with 8𝐵𝑊
represented by (OT-8, IFO-8, CRT-0, SH-0). In this way, we avoid the redundant data rate among

𝐵𝑊…. For IFO-10,

g. When 𝑘 is

𝐵𝑊, 12
212

𝐵𝑊, 8
212

28

different chirps with different OTs, resulting total 23 data rate configurations.

Selective Energy Peak Searching Decoding Method: According to the set of available initial

frequency offsets with the configuration (OT-𝑘, IFO-𝑖, CRT-0, SH-0), we apply the standard dechirp

for decoding and search the energy peak at specific FFT bins corresponding to the available initial

frequency offsets. Since noises are random across the entire spectrum, the fewer the FFT bins

we search for the energy peak, the less the energy peak of noises can influence our energy peak

searching. Thus, when we limit the range of the searching space, we can lower the SNR threshold

for successful decoding. Hence, we provide the tradeoff between noise tolerance and data rate.

6.6.2

Implementation and Evaluation

Setup: We collect a full set of standard LoRa chirp symbols with all available initial frequency

offsets across SF-7 to SF-12 at high SNR levels (>30 dB) on our campus testbed. Then, we generate

chirp symbols at multiple low SNR levels by injecting random Gaussian noise to simulate real-world

scenarios [100]. We use the SNR threshold and data rate as metrics to compare the performance

of different settings. The encoding configuration (OT-𝑘, IFO-𝑖, CRT-0, SH-0) is indicate as SF𝑘-𝑖.

For example, SF7-6 indicates using SF-7 chirps to encode 6-bit data.

Moreover, to verify the battery life gain (BLG) of ChirpTransformer compared to other meth-

ods, we use real-world LoSee traces [95] for vehicle tracking in an urban environment. A mobile

LoRa node on a car periodically travels across the roads between office and home, ranging from

hundreds of meters to 3.2 kilometers in an urban area. The LoRa node transmitted packets with SF-

12 configuration. With the idea of data rate adaption, we reassign the data rate according to the SNR

to reduce transmission time and shorten energy consumption. One baseline method is the default

148

Figure 6.19 SNR threshold and data rate under various configurations for fine-grained data rate
adaption.

LoRa data rate adaptation, which uses SF-7 to SF-11 configurations to replace SF-12 in some cases

with better SNR conditions. However, ChirpTransformer offers a more precise data rate adaptation

scheme with 23 configurations. We estimate the battery life assuming that the LoRa node is pow-

ered by two AA batteries (2200mAh) with 3.6 V voltage and sending packets of different lengths

(from 10 to 30 bytes) 240 times per day, as the LoRa node would do in the morning and evening

vehicle tracking. Based energy profiling on Figure 6.7, we calculate the energy consumption during

two consecutive transmission beginning times by multiplying the currents across different phases

with input voltages. The extra energy consumption caused by ChangeChannelFhss interruption

on the node is neglectable. Then we accumulate the energy over time and calculate the battery life.

Results for SNR Threshold v.s. Data Rate: Figure 6.19 displays the SNR threshold and data

rate in ChirpTransformer with various encoder configurations. The orange bars (e.g., SF-7, SF-8,

SF-9, SF-10, SF-11, and SF-12) represent the SNR threshold of standard LoRa, while the blue

bars show the complementary SNR threshold of ChirpTransformer. As the data rate (e.g., brown
lines) gradually decreases from 7(cid:1)𝐵𝑊
27

212 , the SNR threshold slowly decreases from -6.2 dB
to -25.7 dB. The mean difference between adjacent data rate schemes is about 0.89 dB. Compared

to 6(cid:1)𝐵𝑊

with the standard LoRa, we can see that ChirpTransformer provides a more flexible data rate scheme

with a more diverse SNR threshold.

Results for Battery Life Gain (BLG): Figure 6.20 shows that LoRa data rate adaptation averagely

149

02000400060008000Data Rate (bits/s)SF7SF7-6SF7-5SF8SF8-7SF8-6SF8-5SF9SF9-8SF9-7SF9-6SF10SF10-8SF10-7SF10-6SF11SF11-9SF11-8SF11-7SF12SF12-10SF12-8SF12-6-30-25-20-15-10-5SNR Threshold(dB)ChirpTransformerStandard LoRaData RateFigure 6.20 Battery life measurement with LoSee vehicle tracking traces.

extends battery life by 2.26(cid:2) compared to the fixed-SF-12. On average, ChirpTransformer data

rate adaption provides a battery life of 3.93(cid:2) that of LoRaWAN with fixed-SF-12 and 1.74(cid:2) that

of LoRa data rate adaptation. The reason for this improvement is that ChirpTransformer reassigns

a better data rate for the positions where the measured SNR is between the SNR threshold values

of two standard LoRa SF configurations, whereas conventional rate adaptation in LoRa wastes the

BLG improvement space. Moreover, the BLG trends remain stable as payload size increases.

6.7 Related Work

LoRa Reliable Decoding: By utilizing either multiple gateways and LoRa nodes, recent stud-

ies [155, 233–235, 241, 251] bring extra SNR gains for LoRa transmissions. Charm [155] coor-

dinates multiple gateways to decode weak signals undecodable at any gateway by detecting the

combined energy peak in the spectrum. Choir [241] exploits the correlation across co-located

LoRa nodes, enabling a larger communication range than an individual one. NELoRa [100] devel-

ops a DNN decoder capturing multi-dimension features to obtain 1.84-2.35 dB SNR gains. Osti-

nato [251] uses repeated SF-12 chirps to achieve weak signal decoding beyond SF-12 configuration.

XCopy [235] enhances signal strength by coherently combining retransmitted packets over weak

links to boost weak signal decoding. In contrast, instead of developing a decoder at the gateway and

server side, we obtain SNR gains from the encoder side. ChirpTransformer is parallel with these

works.

150

101418222630Payload Size (bytes)0200400600800Battery Life (days)Fixed-SF12ChirpTransformerLoRaPacket Collision Resolving: Previous research has focused on identifying collisions in the time

or frequency domain. For instance, Choir [241] matches bits to each LoRa node by detecting the

frequency changes caused by oscillator deficiencies. FTrack [240] identifies collisions by exploring

distinct tracks on the spectrum and symbol edges in the time domain. Due to the near-far problem,

existing work cannot implement concurrent transmission with high SIR among different devices in a

large-scale deployment. CIC [229] decodes signals by combining the spectrum from different parts

of a single symbol to cancel interference signals at gateways and support concurrent transmission.

CurvingLoRa [171] enables efficient concurrent transmission with non-linear chirps, but it is not

compatible with COTS LoRa nodes. LMAC [252] utilizes Channel Activity Detection (CAD) to

implement carrier-sense multiple access protocol to avoid collision. Compared to these works, we

propose an orthogonal encoding space at the encoder side to enable concurrent transmission, which

can be realized on COTS LoRa nodes. The decoder design (e.g., CIC, Choir, FTrack) can be further

combined with ChirpTransformer to further improve network performance.

Rate Adaptation in LoRa: To achieve an adaptive data rate [253], LoRa adjusts the data rate by

using various SFs based on received SNR levels. FLoRa [254] focuses on dynamically managing

link parameters to improve network scalability and energy efficiency. DyLoRa [243] establishes

an energy model that associates link properties with transmission parameters. AdapLoRa [244] is

another approach that periodically adjusts resource allocation based on a linear regression process

estimating network lifetime. Current works rely on the standard LoRa encoding scheme and are

unable to implement precise data rates necessary to adapt to diverse environments to extend battery

life. Beyond the standard LoRa, ChirpTransformer provides fine-grained data rate adaption by

selecting different sets of initial frequency to encode data bits with different on-air times.

6.8 Discussion

6.8.1 System Co-existence

ChirpTransformer and the standard LoRa encoder share the same group of basic chirps (e.g.,

SF-7 to SF-12), raising the concern of the co-existence issues when we apply ChirpTransformer

in existing LoRa deployments. For example, when we utilize SH-[9-12] encoder to enhance com-

151

munication reliability, a ChirpTransformer symbol with the pattern of 8 repeated SF-9 chirps may

interfere with other transmissions with SF-9 setting in co-existent standard LoRa deployment.

To measure the co-existence issues, we conduct experiments to emulate real-world signal col-

lision between ChirpTransformer and the standard LoRa signals. First, we generate a ChirpTrans-

former symbol by selecting from one of three SH settings (i.e., SH-[7-10], SH-[8-11], SH-[9-12]).

Then, the ChirpTransformer symbol is superposed with several standard LoRa symbols, charac-

terized by one of the four different SF settings in the corresponding SH setting, with a random

initial frequency offset. The on-air time of these standard LoRa symbols is the same as the Chirp-

Transformer symbol. We set a random time offset between the two superposed signals ranging

from 0 to 1 of ChirpTransformer symbol on-air time. Moreover, the signal-to-interference ratio

(SIR) [171, 247] of standard LoRa to ChirpTransformer indicates the signal strength of standard

LoRa signals compared to ChirpTransformer signals. The higher the SIR is, the stronger the stan-

dard LoRa signals are. The lower the SIR is, the stronger the ChirpTransformer signals are. We

divide the SIR values into 6 ranges: [-30, -20], [-20, -10], [-10, 0], [0, 10], [10, 20], and [20, 30].

With these emulation settings, we mimic the diverse collision conditions of ChirpTransformer and

standard LoRa signals in various real-world scenarios. Finally, we apply our neural-enhanced de-

coder and the standard dechirp to decode ChirpTransformer and standard LoRa symbols from the

collided signals. In each SIR range, we uniformly generate 1,000 collided signals across different

SIR settings, SH settings in ChirpTransformer, and SF settings in the standard LoRa to calculate

the average SER of decoding ChirpTransformer and standard LoRa symbols.

The impact of ChirpTransformer to existing LoRa deployments. As illustrated in Figure 6.21,

it is evident that standard LoRa signals, when stronger than ChirpTransformer signals in the SIR

ranges of [0,10], [10,20], and [20,30], can maintain a SER below 5%. However, when ChirpTrans-

former signals are stronger than standard LoRa signals as [-10,0] SIR, the SER of standard LoRa

signals increases to 27.2%. This increase in SER is more pronounced as the SIR range decreases,

highlighting interference from much stronger ChirpTransformer signals on existing LoRa.

The impact of the standard LoRa to ChirpTransformer. In Figure 6.21, we can also observe

152

Figure 6.21 Concurrent decoding SER of collided ChirpTransformer and standard LoRa signals
under various SIR settings.

that when standard LoRa signals are weaker than ChirpTransformer signals when SIR ranges are

[-30,-20] and [-20,-10], the SER of decoding ChirpTransformer signals remains less than 5%. The

SER rises up to about 20% at the SIR range [-10,0]. When standard LoRa signals are stronger than

ChirpTransformer, the SER of ChirpTransformer increases quickly to 73% at the SIR range [20,30].

Remarks: The results imply a limitation of applying ChirpTransformer in practice that the co-

existence issues are non-negligible. We could have three ways to alleviate the co-existence issues.

Firstly, we can leverage the near-far effect. Since our SH encoder is more noise-resilient than the

standard LoRa for weak signal decoding, ChirpTransformer is primarily adopted by LoRa nodes

located beyond the reach of standard LoRa gateways. This ensures the high SIR of standard LoRa

to ChirpTransformer and standard LoRa signals can be successfully decoded. Secondly, we can

leverage LoRa carrier sense [55, 252] to avoid packet collision with transmission backoff. Thirdly,

we can develop a new collision resolving method at the decoding side to enable successful decoding

under low SIR conditions by borrowing the features from different LoRa packets [229, 238, 240].

For example, using multiple sub-symbol temporal windows of varying lengths [229] allows solv-

ing collision of ChirpTransformer and standard LoRa packets. By merging the spectral results

from these windows, we retain the consistently appearing desired packet and eliminate intermittent

interfering frequency peaks.

153

[-30,-20][-20,-10][-10,0][0,10][10,20][20,30]SIR Range (dB)00.20.40.60.8Symbol Error RateStandard LoRaChirpTransformer6.8.2 Eﬀicient Adaptive Date Rate

ChirpTransformer supports a fine-grained adaptive data rate (ADR) with 23 configurations,

allowing a LoRa nodes to adapt to its deployment scenario optimally. On the other hand, we need

an efficient protocol to enable agile and accurate ADR with low control overhead. A feasible method

is to inherit the current LoRaWAN ADR framework [21,253,255] in MAC commands. Specifically,

LoRa gateways and network servers run an ADR algorithm that adjusts the encoding configuration

of a LoRa node according to the SNR level of recently received packets. The LoRa node will open

a receiving window after a packet transmission to receive the potential configuration request from

LoRa gateways and network servers. When an encoding configuration change is necessitated, the

LoRa gateways will initiate a request for data rate adjustment using LinkADRReq in the receiving

window of the LoRa node. Then, the LoRa node will change its data rate correspondingly and

respond with LinkADRAns for ADR acknowledgment.

By adopting our fine-grained ADR in this framework, a LoRa node will receive more requests

from LoRa gateways and network servers and transmit more ADR acknowledgments. However,

since the LoRa node keeps opening its receiving window after each packet transmission no matter

whether a feedback packet is coming or not, receiving more requests will not bring extra energy

consumption. Moreover, an ADR acknowledgment only contains two bytes of frame payload [253,

255], which is much shorter than a normal LoRa packet. Thus, the extra energy cost of sending

more ADR acknowledgments is affordable regarding the benefits of our fine-grained SDR.

6.9 Conclusion

In this paper, we propose ChirpTransformer, a versatile LoRa encoding framework to enable

the reliable connection between a LoRa node and gateways at extremely low SNR, achieve high

network throughput with collision resolving, and improve energy efficiency in complex environ-

ments. Instead of using only one parameter, SF, to adjust the encoding method in the standard LoRa,

ChirpTransformer develop four chirp features: 1) on-air time of a symbol; 1) available initial fre-

quency offsets; 2) intra-symbol chirp repeating; 3) inter-symbol chirp pattern hopping, to enlarge

the encoding feature space so that we can design different encoders to meet various LoRa deploy-

154

ment requirements. Different encoding methods can be represented with a four-factor tuple (OT,

IFO, CRT, SH). Specifically, we have designed a symbol hopping based encoder for weak signal

decoding, a chirp repeating based encoder for collision resolving, and a selective initial frequency

offset based encoder for fine-grained data rate adaption. We implement ChirpTransformer with

COTS LoRa nodes and SDR. Then, we conduct extensive experiments in both a campus testbed

and real-world trace-driven studies to evaluate the performance of ChirpTransformer. The results

show a 2.38 (cid:2) network coverage, 3.14 (cid:2) network throughput, and 3.93 (cid:2) battery lifetime compared

with the standard LoRa.

155

CHAPTER 7

CONCLUSION

This dissertation systematically addresses critical challenges in deploying low-power, low-cost, and

scalable LoRa-based wireless networking solutions for rural IoT. By developing practical tech-

niques for robust satellite backhaul, improving signal coverage in complex rural environments,

enabling scalable battery-free communication through innovative backscatter methods, and intro-

ducing flexible encoding mechanisms, my work significantly enhances the reliability, scalability,

and efficiency of rural IoT deployments.

To build reliable and scalable LoRa networks for rural IoT, future research must go beyond local

improvements and tackle full-system challenges in real-world conditions. The next step is to design

systems that are more adaptive, intelligent, and well-integrated. First, space- and aerial vehicle-

assisted networks can greatly improve coverage and flexibility but bring new challenges, such as

handling intermittent links, supporting mobile gateways, and managing dynamic links. Second,

embedding integrated sensing and communication into LoRa networks opens opportunities for joint

environment monitoring and signal optimization. It requires new waveform designs and lightweight

protocols with limited IoT resources. Third, to deal with the diversity of rural applications, we

need cross-layer designs that coordinate the physical layer, network control, and application needs

to dynamically balance throughput, energy, and coverage. These open the door to a new generation

of LoRa systems that are not only technically robust but also deployable at scale across complex,

underserved regions.

156

BIBLIOGRAPHY

C. Li and Z. Cao, “Lora networking techniques for large-scale and long-term iot: A down-
to-top survey,” ACM Computing Surveys, vol. 55, no. 3, 2022.

A. Pagano, D. Croce, I. Tinnirello, and G. Vitale, “A survey on lora for smart agriculture:
Current trends and future perspectives,” IEEE Internet of Things Journal, vol. 10, no. 4, pp.
3664–3679, 2023.

C. Li and Z. Cao, “Lora networking techniques for large-scale and long-term iot: A down-
to-top survey,” ACM Computing Surveys (CSUR), 2022.

J. C. Liando, A. Gamage, A. W. Tengourtius, and M. Li, “Known and unknown facts of lora:
Experiences from a large-scale measurement study,” ACM Transactions on Sensor Networks,
vol. 15, no. 2, pp. 1–35, 2019.

C. Li, Y. Ren, S. Tong, S. I. Siam, M. Zhang, J. Wang, Y. Liu, and Z. Cao, “Chirptransformer:
Versatile lora encoding for low-power wide-area iot,” in Proceedings of ACM MobiSys, 2024.

A. Research, “Nb-iot and lte-m issues to boost lora and sigfox near and long-term lead in
lpwa network connections,” 2021.

Z. Sun, H. Yang, K. Liu, Z. Yin, Z. Li, and W. Xu, “Recent advances in lora: A comprehen-
sive survey,” ACM Transactions on Sensor Networks, vol. 18, no. 4, pp. 1–44, 2022.

Y. Ren, L. Liu, C. Li, Z. Cao, and S. Chen, “Is lorawan really wide? fine-grained lora link-
level measurement in an urban environment,” in Proceedings of IEEE ICNP, 2022.

Semtech,
lora-connect/sx1262, accessed 06-Jun-2025.

“Product details

sx1262,”

https://www.semtech.com/products/wireless-rf/

[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10] Y. Ren, A. Gamage, L. Liu, M. Li, S. Chen, Y. Dong, and Z. Cao, “Sateriot: High-
performance ground-space networking for rural iot,” in Proceedings of ACM MobiCom,
2024.

[11] Y. Ren, W. Sun, J. Du, H. Zeng, Y. Dong, M. Zhang, S. Chen, Y. Liu, T. Li, and Z. Cao,
“Demeter: Reliable cross-soil lpwan with low-cost signal polarization alignment,” in Pro-
ceedings of ACM MobiCom, 2024.

[12] Y. Ren, P. Cai, J. Jiang, J. Du, and Z. Cao, “Prism: High-throughput lora backscatter with

non-linear chirps,” in Proceedings of IEEE INFOCOM, 2023.

[13] Y. Ren, G. Li, Y. Liu, Y. Dong, and Z. Cao, “Aeroecho: Towards agricultural low-power
wide-area backscatter with aerial excitation source,” in Proceedins of IEEE INFOCOM,
2025.

[14] A. Pagano, D. Croce, I. Tinnirello, and G. Vitale, “A survey on lora for smart agriculture:
Current trends and future perspectives,” IEEE Internet of Things Journal, vol. 10, no. 4, pp.
3664–3679, 2023.

157

[15] M. Gan, Y. Liu, L. Liu, C. Wu, Y. Dong, H. Zeng, and Z. Cao, “Poster: mmleaf: Versatile
leaf wetness detection via mmwave sensing,” in Proceedings of ACM MobiSys, 2023.

[16] S. Fang, L. Da Xu, Y. Zhu, J. Ahati, H. Pei, J. Yan, and Z. Liu, “An integrated system
for regional environmental monitoring and management based on internet of things,” IEEE
Transactions on Industrial Informatics, vol. 10, no. 2, pp. 1596–1605, 2014.

[17] M. T. Lazarescu, “Design of a wsn platform for long-term environmental monitoring for iot
applications,” IEEE Journal on emerging and selected topics in circuits and systems, vol. 3,
no. 1, pp. 45–54, 2013.

[18] R. Wang, Y. Liu, and R. Müller, “Detection of passageways in natural foliage using

biomimetic sonar,” Bioinspiration & Biomimetics, vol. 17, no. 5, p. 056009, 2022.

[19] M. Grari, M. Yandouzi, I. Idrissi, M. Boukabous, O. Moussaoui, M. AZIZI, and M. MOUS-
SAOUI, “Using iot and ml for forest fire detection, monitoring, and prediction: a literature
review,” Journal of Theoretical and Applied Information Technology, vol. 100, no. 19, pp.
5445–5461, 2022.

[20] A. Sungheetha, R. Sharma et al., “Real time monitoring and fire detection using internet of
things and cloud based drones,” Journal of Soft Computing Paradigm (JSCP), vol. 2, no. 03,
pp. 168–174, 2020.

[21] L. Alliance, “A technical overview of lora and lorawan,” https://lora-alliance.org/resource_

hub/what-is-lorawan/, Retrieved Jun 06, 2025.

[22] C. Li and Z. Cao, “Lora networking techniques for large-scale and long-term iot: A down-

to-top survey,” ACM Computing Surveys, vol. 55, no. 3, pp. 1–36, 2022.

[23] D. Yang, X. Zhang, X. Huang, L. Shen, J. Huang, X. Chang, and G. Xing, “Understanding
power consumption of nb-iot in the wild: tool and large-scale measurement,” in Proceedings
of ACM MobiCom, 2020.

[24] A. Sørensen, H. Wang, M. J. Remy, N. Kjettrup, R. B. Sørensen, J. J. Nielsen, P. Popovski,
and G. C. Madueño, “Modeling and experimental validation for battery lifetime estimation
in nb-iot and lte-m,” IEEE Internet of Things Journal, vol. 9, no. 12, pp. 9804–9819, 2022.

[25] D. Vasisht, J. Shenoy, and R. Chandra, “L2d2: Low latency distributed downlink for leo

satellites,” in Proceedings of ACM SIGCOMM, 2021.

[26] Y. Li, H. Li, W. Liu, L. Liu, Y. Chen, J. Wu, Q. Wu, J. Liu, and Z. Lai, “A case for stateless
mobile core network functions in space,” in Proceedings of ACM SIGCOMM, 2022.

[27] SpaceX, “Space x,” https://www.spacex.com/, accessed 06-Jun-2025.

[28] SWARM, “Space x swarm -

low cost,

global

satellite connectivity for

iot,”

https://swarm.space/, accessed 06-Jun-2025.

[29] B. R. Elbert, Introduction to satellite communication. Artech house, 2008.

158

[30] G. Maral, M. Bousquet, and Z. Sun, Satellite communications systems: systems, techniques

and technology.

John Wiley & Sons, 2020.

[31] L. J. Ippolito Jr, Satellite communications systems engineering: atmospheric effects, satellite

link design and system performance.

John Wiley & Sons, 2017.

[32] M. Centenaro, C. E. Costa, F. Granelli, C. Sacchi, and L. Vangelista, “A survey on tech-
nologies, standards and open challenges in satellite iot,” IEEE Communications Surveys &
Tutorials, vol. 23, no. 3, pp. 1693–1720, 2021.

[33] L. J. Ippolito, Radiowave propagation in satellite communications.

Springer Science &

Business Media, 2012.

[34] SWARM, “Product overview swarm m138 modem,” https://swarm.space/wp-content/

uploads/2022/09/Swarm-M138-Specifications.pdf, accessed 06-Jun-2025.

[35]

Iridium, “Urb-iot: Lorawan outdoor gateway,” https://www.iridium.com/products/urb-iot/,
accessed 06-Jun-2025.

[36] LONESTAR,

“Lorawan

satellite

gateway,”

https://www.lonestartracking.com/

lorawan-satellite-gateway/, accessed 06-Jun-2025.

[37] DRYAD, “Fighting wildfires with long-range, low-power technologies,” https://https://www.

dryad.net/post/fighting-wildfires-with-lorawan, accessed 06-Jun-2025.

[38] SWARM, “Case study | dryad networks,” https://swarm.space/wp-content/uploads/2022/02/

Swarm-Environmental-Case-Study-Dryad.pdf, accessed 06-Jun-2025.

[39] C. Li, Y. Ren, S. Tong, S. I. Siam, M. Zhang, J. Wang, Y. Liu, and Z. Cao, “Chirptransformer:
Versatile lora encoding for low-power wide-area iot,” in Proceedings of ACM MobiSys, 2024.

[40]

J. Du, Y. Ren, Z. Zhu, C. Li, Z. Cao, Q. Ma, and Y. Liu, “Srlora: Neural-enhanced lora weak
signal decoding with multi-gateway super resolution,” in Proceedings of ACM MobiHoc,
2023.

[41] T. Winter, P. Thubert, A. Brandt, J. Hui, R. Kelsey, P. Levis, K. Pister, R. Struik, J.-P. Vasseur,
and R. Alexander, “Rpl: Ipv6 routing protocol for low-power and lossy networks,” Tech.
Rep., 2012.

[42] O. Gnawali, R. Fonseca, K. Jamieson, D. Moss, and P. Levis, “Collection tree protocol,” in

Proceedings of ACM Sensys, 2009.

[43] F. Zhang, F. Yu, X. Zheng, L. Liu, and H. Ma, “Dfh: Improving the reliability of lr-fhss via

dynamic frequency hopping,” in Proceedings of IEEE ICNP, 2023.

[44] P. I. Parra, S. Montejo-Sánchez, J. A. Fraire, R. D. Souza, and S. Céspedes, “Network size
estimation for direct-to-satellite iot,” IEEE Internet of Things Journal, vol. 10, no. 7, pp.
6111–6125, 2022.

[45] Starlink, “Space x starlink,” https://www.starlink.com/, accessed 06-Jun-2025.

159

[46] OneWeb, “Oneweb constellation,” https://oneweb.net/, accessed 06-Jun-2025.

[47] Sateliot, “Sateliot,” https://sateliot.space/en/, accessed 06-Jun-2025.

[48] Lacuna, “Lacuna space,” https://lacuna.space/, accessed 06-Jun-2025.

[49] Semtech, “Semtech smart agriculture,” https://www.semtech.com/lora/lora-applications/

smart-agriculture, accessed 06-Jun-2025.

[50] SWARM, “Swarm eva kit,” https://swarm.space/product/swarm-eval-kit/, accessed 06-Jun-

2025.

[51] O. Kodheli, E. Lagunas, N. Maturo, S. K. Sharma, B. Shankar, J. F. M. Montoya, J. C. M.
Duncan, D. Spano, S. Chatzinotas, S. Kisseleff et al., “Satellite communications in the new
space era: A survey and future challenges,” IEEE Communications Surveys & Tutorials,
vol. 23, no. 1, pp. 70–109, 2020.

[52] S. Demetri, M. Zúñiga, G. P. Picco, F. Kuipers, L. Bruzzone, and T. Telkamp, “Automated
estimation of link quality for lora: A remote sensing approach,” in Proceedings of ACM/IEEE
IPSN, 2019.

[53] Z. Cao, J. Wang, D. Liu, Q. Ma, X. Miao, and X. Mao, “Chase++: Fountain-enabled fast
flooding in asynchronous duty cycle networks,” IEEE/ACM Transactions on Networking,
vol. 29, no. 1, pp. 410–422, 2020.

[54] Z. Li, M. Li, J. Wang, and Z. Cao, “Ubiquitous data collection for mobile users in wireless

sensor networks,” in Proceedings of IEEE INFOCOM, 2011.

[55] A. Gamage, J. C. Liando, C. Gu, R. Tan, and M. Li, “Lmac: efficient carrier-sense multiple

access for lora,” in Proceedings of ACM MobiCom, 2020.

[56] P. Levis, N. Patel, D. Culler, and S. Shenker, “Trickle: A self-regulating algorithm for code
propagation and maintenance in wireless sensor networks,” in Proceedings of USENIX/ACM
NSDI, 2004.

[57] T.

Connectivity,

“Ant-916-cw-rcs-sma,”

product-ANT-916-CW-RCL-SMA.html?q=CW-RCL&source=header,
Jun-2025.

https://www.te.com/usa-en/
06-

accessed

[58] ESPRESSIF, “Esp32,” https://www.espressif.com/en/products/socs/esp32, accessed 06-

Jun-2025.

[59] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in Proceedings of

ACM SIGKDD, 2016.

[60] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: synthetic minority
over-sampling technique,” Journal of artificial intelligence research, vol. 16, pp. 321–357,
2002.

160

[61] R. Pi, “Raspberry pi 4,” https://www.raspberrypi.com/products/raspberry-pi-4-model-b/,

accessed 06-Jun-2025.

[62] Semtech, “Lora calculator,” https://lora-developers.semtech.com/build/tools/calculator/,

Retrieved Sep 11, 2023.

[63] B. Tao, M. Masood, I. Gupta, and D. Vasisht, “Transmitting, fast and slow: Scheduling

satellite traffic through space and time,” in Proceedings of ACM MobiCom, 2023.

[64] Y. Li, H. Li, W. Liu, L. Liu, W. Zhao, Y. Chen, J. Wu, Q. Wu, J. Liu, Z. Lai et al., “A
networking perspective on starlink’s self-driving leo mega-constellation,” in Proceedings of
ACM MobiCom, 2023.

[65] Z. Lai, H. Li, Y. Deng, Q. Wu, J. Liu, Y. Li, J. Li, L. Liu, W. Liu, and J. Wu, “fStarryNetg:
Empowering researchers to evaluate futuristic integrated space and terrestrial networks,” in
Proceedings of USNEIX NSDI), 2023.

[66] B. Tao, O. Chabra, I. Janveja, I. Gupta, and D. Vasisht, “Known knowns and unknowns:
Near-realtime earth observation via query bifurcation in serval,” in Proceedings of USENIX
NSDI, 2024.

[67]

J. A. Fraire, S. Henn, F. Dovis, R. Garello, and G. Taricco, “Sparse satellite constellation
design for lora-based direct-to-satellite internet of things,” in Proceedings of IEEE GLOBE-
COM, 2020.

[68] Z. Zhang, Y. Li, C. Huang, Q. Guo, L. Liu, C. Yuen, and Y. L. Guan, “User activity detec-
tion and channel estimation for grant-free random access in leo satellite-enabled internet of
things,” IEEE Internet of Things journal, vol. 7, no. 9, pp. 8811–8825, 2020.

[69] Y. Qian, L. Ma, and X. Liang, “Symmetry chirp spread spectrum modulation used in leo
satellite internet of things,” IEEE Communications Letters, vol. 22, no. 11, pp. 2230–2233,
2018.

[70] V. Singh, T. Chakraborty, S. Jog, O. Chabra, D. Vasisht, and R. Chandra, “Spectrumize:
Spectrum-efficient satellite networks for the internet of things,” in Proceedings of USENIX
NSDI, 2024.

[71] D. Vasisht, Z. Kapetanovic, J. Won, X. Jin, R. Chandra, S. Sinha, A. Kapoor, M. Sudarshan,
and S. Stratman, “Farmbeats: An iot platform for data-driven agriculture,” in Proceedings
of USENIX NSDI, 2017.

[72] N. Islam, B. Ray, and F. Pasandideh, “Iot based smart farming: Are the lpwan technologies

suitable for remote communication?” in Proceedings of IEEE SmartIoT, 2020.

[73] R. Gebbers and V. I. Adamchuk, “Precision agriculture and food security,” Science, 2010.

[74] N. Ahmed, D. De, and I. Hussain, “Internet of things (iot) for smart precision agriculture

and farming in rural areas,” IEEE Internet of Things Journal, 2018.

161

[75] EPQ, “Ground water,” https://www.epa.gov/report-environment/ground-water#importance,

retrieved by Aug 15th, 2023.

[76] USEPA, “Estimated nitrate concentrations in groundwater used for drinking,” https://www.
epa.gov/nutrient-policy-data/estimated-nitrate-concentrations-groundwater-used-drinking,
accessed 06-Jun-2025.

[77] ——,

gases,”
overview-greenhouse-gases, accessed 06-Jun-2025.

“Overview of

greenhouse

https://www.epa.gov/ghgemissions/

[78] N. Quintana-Ashwell, D. Gholson, G. Kaur, G. Singh, J. Massey, L. J. Krutz, C. G. Henry,
T. Cooke III, M. Reba, and M. A. Locke, “Irrigation water management tools and alterna-
tive irrigation sources trends and perceptions by farmers from the delta regions of the lower
mississippi river basin in south central usa,” Agronomy, 2022.

[79] C. A. Carlson and J. L. Ingraham, “Comparison of denitrification by pseudomonas stutzeri,
pseudomonas aeruginosa, and paracoccus denitrificans,” Applied and Environmental Micro-
biology, 1983.

[80] D. E. A. Center, “Nitrate in drinking water,” https://www.dhd2.org/wp-content/uploads/

2016/12/deq-wd-gws-ciu-nitratebrochure_270430_7.pdf, accessed 06-Jun-2025.

[81] Doktar,

“Filiz:

Agricultural

sensor

station,”

https://www.doktar.com/en/products/

agricultural-sensor-station-soil-temperature/, accessed 06-Jun-2025.

[82] Monnit,

“Monnit agriculture and livestock monitoring,” https://www.monnit.com/

applications/agriculture-livestock-monitoring/, accessed 06-Jun-2025.

[83] HoBo,

“Hobonet

field

monitoring

system,”

https://www.onsetcomp.com/

hobonet-field-monitoring-system, retrieved by Aug 14th, 2023.

[84]

I. F. Akyildiz and E. P. Stuntebeck, “Wireless underground sensor networks: Research chal-
lenges,” Ad Hoc Networks, 2006.

[85] A. R. Silva and M. C. Vuran, “Cps2: integration of center pivot systems with wireless under-
ground sensor networks for autonomous precision agriculture,” in Proceedings of ACM/IEEE
ICCPS, 2010.

[86] NSF, “2022 signal in the soils program solicitation,” https://www.nsf.gov/pubs/2022/

nsf22550/nsf22550.htm, accessed 06-Jun-2025.

[87] Z. Sun and I. F. Akyildiz, “Connectivity in wireless underground sensor networks,” in Pro-

ceedings of IEEE SECON, 2010.

[88] B. Zhou, V. S. S. L. Karanam, and M. C. Vuran, “Impacts of soil and antenna characteristics

on lora in internet of underground things,” in Proceedings of IEEE GLOBECOM, 2021.

[89] K. Lin and T. Hao, “Experimental link quality analysis for lora-based wireless underground

sensor networks,” IEEE Internet of Things Journal, 2021.

162

[90] Z. Sun and I. F. Akyildiz, “Magnetic induction communications for wireless underground

sensor networks,” IEEE transactions on antennas and propagation, 2010.

[91] A. R. Silva and M. C. Vuran, “Communication with aboveground devices in wireless under-

ground sensor networks: An empirical study,” in Proceedings of IEEE ICC, 2010.

[92] K. Yang, Y. Chen, X. Chen, and W. Du, “Link quality modeling for lora networks in or-

chards,” in Proceedings of ACM/IEEE IPSN, 2023.

[93] A. R. Silva and M. C. Vuran, “Empirical evaluation of wireless underground-to-underground
communication in wireless underground sensor networks,” in Proceedings of DCOSS, 2009.

[94] M. C. Vuran and I. F. Akyildiz, “Channel model and analysis for wireless underground sensor

networks in soil medium,” Physical communication, 2010.

[95] Y. Ren, L. Liu, C. Li, Z. Cao, and S. Chen, “Is lorawan really wide? fine-grained lora link-
level measurement in an urban environment,” in Proceedings of IEEE ICNP, 2022.

[96] M. Mezzavilla, M. Zhang, M. Polese, R. Ford, S. Dutta, S. Rangan, and M. Zorzi, “End-
to-end simulation of 5g mmwave networks,” IEEE Communications Surveys & Tutorials,
2018.

[97] X.-f. Wan, Y. Yang, J. Cui, and M. S. Sardar, “Lora propagation testing in soil for wireless

underground sensor networks,” in Proceedings of IEEE APCAP, 2017.

[98] C. Ebi, F. Schaltegger, A. Rüst, and F. Blumensaat, “Synchronous lora mesh network to

monitor processes in underground infrastructure,” IEEE Access, 2019.

[99] G. Di Renzone, S. Parrino, G. Peruzzi, A. Pozzebon, and D. Bertoni, “Lorawan underground
to aboveground data transmission performances for different soil compositions,” IEEE Trans-
actions on Instrumentation and Measurement, 2021.

[100] C. Li, H. Guo, S. Tong, X. Zeng, Z. Cao, M. Zhang, Q. Yan, L. Xiao, J. Wang, and Y. Liu,
“Nelora: Towards ultra-low snr lora communication with neural-enhanced demodulation,”
in Proceedings of ACM Sensys, 2021.

[101] T. E. Berry, E. Lord, and C. Morgan, “Soil polarization data collected for the global undis-
turbed/disturbed earth (guide) program,” in SPIE Proceedings of Polarization: Measure-
ment, Analysis, and Remote Sensing XII, 2016.

[102] G. Hassan, K. Ismail, N. Persaud, and R. Reneau Jr, “Dependence of the degree of linear
polarization in scattered visible light on soil textural fractions,” Soil science, 2004.

[103] A. Sarkar, “Change of polarization of electromagnetic waves on reflection from planar in-

terfaces,” IETE Journal of Research, 1981.

[104] T. Tanikawa, K. Masuda, H. Ishimoto, T. Aoki, M. Hori, M. Niwano, A. Hachikubo, S. Ma-
toba, K. Sugiura, T. Toyota et al., “Spectral degree of linear polarization and neutral points of
polarization in snow and ice surfaces,” Journal of Quantitative Spectroscopy and Radiative
Transfer, 2021.

163

[105] F.-M. Breon, D. Tanre, P. Lecomte, and M. Herman, “Polarized reflectance of bare soils
and vegetation: measurements and models,” IEEE Transactions on Geoscience and Remote
Sensing, 1995.

[106] Semtech, “Semtech,” https://www.semtech.com/, accessed 06-Jun-2025.

[107] T. A. Milligan, Modern antenna design.

John Wiley & Sons, 2005.

[108] M. Nilsson, “Directional antennas for wireless sensor networks,” in In Proceedings of Adhoc

Workshop, 2009.

[109] B. Großwindhager, M. S. Bakr, M. Rath, F. Gentili, W. Bösch, K. Witrisal, C. A. Boano, and
K. Römer, “Poster: Switchable directional antenna system for uwb-based internet of things
applications.” in Proceedings of EWSN, 2017.

[110] Y. Wang, X. Zheng, L. Liu, and H. Ma, “Polartracker: Attitude-aware channel access for
floating low power wide area networks,” in Proceedings of IEEE INFOCOM, 2021.

[111] R. Li, X. Zheng, Y. Wang, L. Liu, and H. Ma, “Polarscheduler: Dynamic transmission control

for floating lora networks,” in Proceedings of IEEE INFOCOM, 2022.

[112] Semtech, “Sx1302,” https://www.semtech.com/products/wireless-rf/lora-core/sx1302, ac-

cessed 06-Jun-2025.

[113] ——,

“Sx1303,” https://www.semtech.com/products/wireless-rf/lora-core/sx1303,

ac-

cessed 06-Jun-2025.

[114] ——,

“sx1302cfd915gw1h,”

https://www.semtech.com/products/wireless-rf/lora-core/

sx1302cfd915gw1h, accessed 06-Jun-2025.

[115] L. Chen, W. Hu, K. Jamieson, X. Chen, D. Fang, and J. Gummeson, “Pushing the physical
limits of iot devices with programmable metasurfaces,” in Proceedings of USENIX NSDI,
2021.

[116] J. D. van der Laan, J. B. Wright, S. A. Kemme, and D. A. Scrymgeour, “Superior signal per-
sistence of circularly polarized light in polydisperse, real-world fog environments,” Applied
Optics, 2018.

[117] J. Van der Laan, D. Scrymgeour, S. Kemme, and E. Dereniak, “Detection range enhancement
using circularly polarized light in scattering environments for infrared wavelengths,” Applied
optics, 2015.

[118] B. Y. Toh, R. Cahill, and V. F. Fusco, “Understanding and measuring circular polarization,”

IEEE Transactions on Education, 2003.

[119] M. Nicolet, M. Schnaiter, and O. Stetzer, “Circular depolarization ratios of single water
droplets and finite ice circular cylinders: a modeling study,” Atmospheric Chemistry and
Physics, 2012.

164

[120] A. Ishimaru, Electromagnetic wave propagation, radiation, and scattering: from fundamen-

tals to applications.

John Wiley & Sons, 2017.

[121] J. D. van der Laan, J. B. Wright, D. A. Scrymgeour, S. A. Kemme, and E. L. Dereniak,
“Evolution of circular and linear polarization in scattering environments,” Optics Express,
2015.

[122] V. Temnov, K. Sokolowski-Tinten, P. Zhou, A. El-Khamhawy, and D. Von Der Linde, “Mul-
tiphoton ionization in dielectrics: comparison of circular and linear polarization,” Physical
review letters, 2006.

[123] R. Galuscak-OM6AA and P. Hazdra, “Circular polarization and polarization losses,” 2006.

[124] L. Mottola, T. Voigt, and G. P. Picco, “Electronically-switched directional antennas for wire-
less sensor networks: A full-stack evaluation,” in Proceedings of IEEE SECON, 2013.

[125] M. Nilsson, “Spida: A direction-finding antenna for wireless sensor networks,” in Proceed-

ings of Springer International Workshop on Real-World Wireless Sensor Networks, 2010.

[126] G. Tarter, L. Mottola, and G. P. Picco, “Directional antennas for convergecast in wireless

sensor networks: Are they a good idea?” in Proceedings of IEEE MASS, 2016.

[127] J. Schandy, S. Olofsson, N. Gammarano, L. Steinfeld, and T. Voigt, “Improving sensor net-
work performance with directional antennas: A cross-layer optimization,” ACM Transactions
on Sensor Networks, 2021.

[128] N. G. Apergis, I. P. Inglesis, S. Vasileiadis, and S. A. Mitilineos, “Development of a small
form factor switched-beam antenna for iot applications at the 868 mhz ism band,” in Pro-
ceedings of APC, 2019.

[129] M. A. R. Hashmi and P. V. Brennan, “An antenna solution for glacial environmental sensor

networks,” IEEE Access, 2023.

[130] Y. Zhang, W. Sun, Y. Ren, S.-j. Lee, and M. Li, “Channel adapted antenna augmentation for

improved wi-fi throughput,” IEEE Transactions on Mobile Computing, 2022.

[131] pSemi,

“pe44820,”

https://www.psemi.com/products/rf-phase-amplitude-control/

phase-shifters/pe44820, accessed 06-Jun-2025.

[132] XRDS-RF, “3db xrds-rf 698-2700mhz passive splitter,” https://www.amazon.com/
accessed 06-

Wide-Band-Splitter-3dB-Type-Female-50-XRDS-RF/dp/B07V6YH8BV,
Jun-2025.

[133] WAVEFORM, “2x2 cross-polarized 600-2700 mhz outdoor antenna,” https://www.
waveform.com/products/2x2-mimo-outdoor-panel-antenna, accessed 06-Jun-2025.

[134] Semtech, “Sx1276 datasheet,” https://www.semtech.com/products/wireless-rf/lora-connect/

sx1276, accessed 06-Jun-2025.

165

[135] ——,

“Lora

and

lorawan,”

https://lora-developers.semtech.com/documentation/

tech-papers-and-guides/lora-and-lorawan/, accessed 06-Jun-2025.

[136] S. Tong, J. Wang, J. Yang, Y. Liu, and J. Zhang, “Citywide lora network deployment and
operation: Measurements, analysis, and implications,” in Proceedings of ACM SenSys, 2023.

[137] USDA, “Soil texture calculator,” Retrieved by Dec 27th 2023.

[138] D. G. Groenendyk, T. P. Ferre, K. R. Thorp, and A. K. Rice, “Hydrologic-process-based soil

texture classifications for improved visualization of landscape function,” PloS one, 2015.

[139] USDA, “Web soil survey,” Retrieved by Dec 27th 2023.

[140] PINO-TECH, “Soilwatch 10 – soil moisture sensor,” https://pino-tech.eu/soilwatch10/, ac-

cessed 06-Jun-2025.

[141] S. Balivada, G. Grant, X. Zhang, M. Ghosh, S. Guha, and R. Matamala, “A wireless under-
ground sensor network field pilot for agriculture and ecology: Soil moisture mapping using
signal attenuation,” Sensors, 2022.

[142] N. Chaamwe, W. Liu, and H. Jiang, “Wave propagation communication models for wireless

underground sensor networks,” in Proceedings of IEEE ICCT, 2010.

[143] C. L. Holloway, E. F. Kuester, J. A. Gordon, J. O’Hara, J. Booth, and D. R. Smith, “An
overview of the theory and applications of metasurfaces: The two-dimensional equivalents
of metamaterials,” IEEE antennas and propagation magazine, 2012.

[144] S. Okamura and T. Oguchi, “Electromagnetic wave propagation in rain and polarization ef-

fects,” Proceedings of the Japan academy, series B, 2010.

[145] K. Y. Bliokh and Y. P. Bliokh, “Conservation of angular momentum, transverse shift, and
spin hall effect in reflection and refraction of an electromagnetic wave packet,” Physical
review letters, 2006.

[146] J. A. Shaw, “Degree of linear polarization in spectral radiances from water-viewing infrared

radiometers,” Applied optics, 1999.

[147] J. A. Stratton, Electromagnetic theory.

John Wiley & Sons, 2007, vol. 33.

[148] D. M. Pozar, Microwave engineering.

John wiley & sons, 2011.

[149] C.-I. Shie, J.-C. Cheng, S.-C. Chou, and Y.-C. Chiang, “Transdirectional coupled-line cou-
plers implemented by periodical shunt capacitors,” IEEE Transactions on Microwave Theory
and Techniques, 2009.

[150] A. Salam and U. Raza, Signals in the Soil. Springer, 2020.

[151] M. C. Vuran and A. R. Silva, “Communication through soil in wireless underground sensor
networks–theory and practice,” Sensor networks: Where theory meets practice, 2009.

166

[152] E. Research, “Usrp n210 datasheet,” https://www.ettus.com/all-products/un210-kit/, ac-

cessed 06-Jun-2025.

[153] ——, “Sbx 400-4400 datasheet,” https://www.ettus.com/all-products/sbx/, accessed 06-Jun-

2025.

[154] J. Du, Y. Ren, M. Zhang, Y. Liu, and Z. Cao, “Nelora-bench: A benchmark for
neural-enhanced lora demodulation,” International Conference on Learning Representations
(ICLR) Workshop on Machine Learning for IoT, 2023.

[155] A. Dongare, R. Narayanan, A. Gadre, A. Luong, A. Balanuta, S. Kumar, B. Iannucci, and
A. Rowe, “Charm: Exploiting geographical diversity through coherent combining in low-
power wide-area networks,” in Proceedings of ACM/IEEE IPSN, 2018.

[156] Q. Wang, A. Terzis, and A. Szalay, “A novel soil measuring wireless sensor network,” in

Proceedings of IEEE I2MTC, 2010.

[157] J. Tooker, X. Dong, M. C. Vuran, and S. Irmak, “Connecting soil to the cloud: A wireless

underground sensor network testbed,” in Proceedings of IEEE SECON, 2012.

[158] X. Dong, M. C. Vuran, and S. Irmak, “Autonomous precision agriculture through integra-
tion of wireless underground sensor networks with center pivot irrigation systems,” Ad Hoc
Networks, 2013.

[159] G. Castellanos, M. Deruyck, L. Martens, and W. Joseph, “System assessment of wusn using

nb-iot uav-aided networks in potato crops,” IEEE Access, 2020.

[160] S. Yang, O. Baltaji, A. C. Singer, and Y. M. Hashash, “Development of an underground
through-soil wireless acoustic communication system,” IEEE Wireless Communications,
2019.

[161] J. Vasquez, V. Rodriguez, and D. Reagor, “Underground wireless communications using
high-temperature superconducting receivers,” IEEE Transactions on Applied Superconduc-
tivity, 2004.

[162] Q. Zhu and H. Lin, “Influences of soil, terrain, and crop growth on soil moisture variation

from transect to farm scales,” Geoderma, 2011.

[163] S. Datta, S. Taghvaeian, and J. Stivers, “Understanding soil water content and thresholds for
irrigation management,” Oklahoma Cooperative Extension Service, Tech. Rep., 2017.

[164] M. E. Holzman, R. Rivas, and M. C. Piccolo, “Estimating soil moisture and the relationship
with crop yield using surface temperature and vegetation index,” International Journal of
Applied Earth Observation and Geoinformation, 2014.

[165] C. Nendel, M. Berg, K. C. Kersebaum, W. Mirschel, X. Specka, M. Wegehenkel, K. Wenkel,
and R. Wieland, “The monica model: Testing predictability for crop growth, soil moisture
and nitrogen dynamics,” Ecological Modelling, 2011.

167

[166] Y. Peng, L. Shangguan, Y. Hu, Y. Qian, X. Lin, X. Chen, D. Fang, and K. Jamieson, “Plora:
A passive long-range data network from ambient lora transmissions,” in Proceedings of ACM
SIGCOMM, 2018.

[167] V. Talla, M. Hessar, B. Kellogg, A. Najafi, J. R. Smith, and S. Gollakota, “Lora backscatter:
Enabling the vision of ubiquitous connectivity,” ACM IMWUT, vol. 1, no. 3, 2017.

[168] J. Jiang, Z. Xu, F. Dang, and J. Wang, “Long-range ambient lora backscatter with parallel

decoding,” in Proceedings of ACM MobiCom, 2021.

[169] M. Hessar, A. Najafi, and S. Gollakota, “Netscatter: Enabling large-scale backscatter net-

works,” in Proceedings of USENIX NSDI, 2019.

[170] W. Mao, Z. Zhao, Z. Chang, G. Min, and W. Gao, “Energy-efficient industrial internet of
things: overview and open issues,” IEEE transactions on industrial informatics, vol. 17,
no. 11, pp. 7225–7237, 2021.

[171] C. Li, X. Guo, L. Shangguan, Z. Cao, and K. Jamieson, “Curvinglora to boost lora network

throughput via concurrent transmission,” in Proceedings of USENIX NSDI, 2022.

[172] C. Li, Z. Cao, and L. Xiao, “Curvealoha: Non-linear chirps enabled high throughput random

channel access for lora,” in Proceedings of IEEE INFOCOM, 2022.

[173] X. Guo, L. Shangguan, Y. He, J. Zhang, H. Jiang, A. A. Siddiqi, and Y. Liu, “Aloba: rethink-
ing on-off keying modulation for ambient lora backscatter,” in Proceedings of ACM SenSys,
2020.

[174] Avago, “Hsms-285c datasheet,” https://docs.broadcom.com/doc/AV02-1377EN, accessed

06-Jun-2025.

[175] T.

Instruments, “Lpv7215mg datasheet,” https://www.ti.com/store/ti/en/p/product/?p=

LPV7215MG/NOPB, accessed 06-Jun-2025.

[176] STMicroelectronics,

“Stm32l011d3p6

datasheet,”

https://www.st.com/resource/en/

datasheet/stm32l011d4.pdf, accessed 06-Jun-2025.

[177] Maxim,

“Max5530

datasheet,”

https://datasheets.maximintegrated.com/en/ds/

MAX5530-MAX5531.pdf, accessed 06-Jun-2025.

[178] A.

Devices,

“Ltc6990is6

datasheet,”

https://www.analog.com/media/en/

technicaldocumentation/data-sheets/LTC6990.pdf, accessed 06-Jun-2025.

[179] ——, “Adg902 datasheet,” https://www.analog.com/media/en/technicaldocumentation/

data-sheets/ADG901_902.pdf, accessed 06-Jun-2025.

[180] C. Li, H. Guo, S. Tong, X. Zeng, Z. Cao, M. Zhang, Q. Yan, L. Xiao, J. Wang, and Y. Liu,
“Nelora: Towards ultra-low snr lora communication with neural-enhanced demodulation,”
in Proceedings of ACM SenSys, 2021.

168

[181] V. Liu, A. Parks, V. Talla, S. Gollakota, D. Wetherall, and J. R. Smith, “Ambient backscatter:
Wireless communication out of thin air,” ACM SIGCOMM computer communication review,
vol. 43, no. 4, p. 39–50, 2013.

[182] R. Zhao, F. Zhu, Y. Feng, S. Peng, X. Tian, H. Yu, and X. Wang, “Ofdma-enabled wi-fi

backscatter,” in Proceedings of ACM MobiCom, 2019.

[183] Y. K. Chan, M. Y. Chua, and V. Koo, “Sidelobes reduction using simple two and tri-stages
non linear frequency modulation (nlfm),” Progress In Electromagnetics Research, vol. 98,
pp. 33–52, 2009.

[184] S. Boukeffa, Y. Jiang, and T. Jiang, “Sidelobe reduction with nonlinear frequency modulated
waveforms,” in IEEE International Colloquium on Signal Processing and its Applications,
2011.

[185] N. Abramson, “The aloha system: Another alternative for computer communications,” in
Proceedings of the November 17-19, 1970, fall joint computer conference, 1970, pp. 281–
285.

[186] R. Eletreby, D. Zhang, S. Kumar, and O. Yağan, “Empowering low-power wide area networks

in urban settings,” in Proceedings of ACM SIGCOMM, 2017.

[187] X. Xia, Y. Zheng, and T. Gu, “Ftrack: Parallel decoding for lora transmissions,” IEEE/ACM

Transactions on Networking, vol. 28, no. 6, pp. 2573–2586, 2020.

[188] M. O. Shahid, M. Philipose, K. Chintalapudi, S. Banerjee, and B. Krishnaswamy, “Concur-
rent interference cancellation: Decoding multi-packet collisions in lora,” in Proceedings of
ACM SIGCOMM, 2021.

[189] X. Wang, L. Kong, L. He, and G. Chen, “mLoRa: A Multi-Packet Reception Protocol in

LoRa networks,” in Proceedings of IEEE ICNP, 2019.

[190] S. Tong, Z. Xu, and J. Wang, “Colora: Enabling multi-packet reception in lora,” in Proceed-

ings of IEEE INFOCOM, 2020.

[191] Z. Xu, P. Xie, and J. Wang, “Pyramid: Real-time lora collision decoding with peak tracking,”

in Proceedings of IEEE INFOCOM, 2021.

[192] S. Tong, J. Wang, and Y. Liu, “Combating packet collisions using non-stationary signal scal-

ing in lpwans,” in Proceedings of ACM MobiSys, 2020.

[193] X. Xia, N. Hou, Y. Zheng, and T. Gu, “Pcube: scaling lora concurrent transmissions with

reception diversities,” in Proceedings of ACM MobiCom, 2021.

[194] A. Gamage, J. Liando, C. Gu, R. Tan, M. Li, and O. Seller, “Lmac: Efficient carrier-sense
multiple access for lora,” ACM Transactions on Sensor Networks, vol. 19, no. 2, pp. 1–27,
2023.

[195] N. Kouvelas, V. S. Rao, R. V. Prasad, G. Tawde, and K. Langendoen, “P-carma: Politely

scaling lorawan,” in Proceedings of ACM MobiCom, 2020.

169

[196] O. Elijah, T. A. Rahman, I. Orikumhi, C. Y. Leow, and M. N. Hindia, “An overview of
internet of things (iot) and data analytics in agriculture: Benefits and challenges,” IEEE IoT
Journal, vol. 5, no. 5, pp. 3758–3773, 2018.

[197] “FarmBeats: AI, Edge, and IoT for Smart Farms,” https://www.microsoft.com/en-us/

research/project/farmbeats-iot-agriculture/, Retrieved by Jul 28th 2023.

[198] M. S. Islam and G. K. Dey, “Precision agriculture: Renewable energy based smart crop field
monitoring and management system using wsn via iot,” in Proceedings of IEEE STI, 2019.

[199] Y. Liu, M. Gan, H. Zeng, L. Liu, Y. Dong, and Z. Cao, “Hydra: Accurate multi-modal
leaf wetness sensing with mm-wave and camera fusion,” in Proceedings of ACM MobiCom,
2024.

[200] T. Chakraborty, H. Shi, Z. Kapetanovic, B. Priyantha, D. Vasisht, B. Vu, P. Pandit, P. Pillai,
Y. Chabria, A. Nelson, M. Daum, and R. Chandra, “Whisper: Iot in the tv white space
spectrum,” in Proceedings of USENIX NSDI, 2022.

[201] Y. Ren, Y. Wang, Y. Dong, S. Chen, M. Zhang, J. Tang, and Z. Cao, “Demeter-demo:
Demonstrating cross-soil lpwan with low-cost signal polarization alignment,” in Proceed-
ings of ACM MobiCom, 2024.

[202] D. Yang, G. Xing, J. Huang, X. Chang, and X. Jiang, “Qid: Robust mobile device recognition
via a multi-coil qi-wireless charging system,” ACM Transactions on Internet of Things, vol. 3,
no. 2, pp. 1–27, 2022.

[203] P. Zhang, P. Hu, V. Pasikanti, and D. Ganesan, “Ekhonet: High speed ultra low-power

backscatter for next generation sensors,” in Proceedings of ACM MobiCom, 2014.

[204] M. Katanbaf, A. Weinand, and V. Talla, “Simplifying backscatter deployment: Full-duplex

lora backscatter,” in Proceedings of USENIX NSDI, 2021.

[205] Y. Ren, P. Cai, J. Jiang, J. Du, and Z. Cao, “Prism: High-throughput lora backscatter with

non-linear chirps,” in Proceedings of IEEE INFOCOM, 2023.

[206] Y. Song, L. Lu, J. Wang, C. Zhang, H. Zheng, S. Yang, J. Han, and J. Li, “𝜇mote: Enabling
passive chirp de-spreading and 𝜇w-level long-range downlink for backscatter devices,” in
Proceedings of USENIX NSDI, 2023.

[207] S. Li, H. Zheng, C. Zhang, Y. Song, S. Yang, M. Chen, L. Lu, and M. Li, “Passive dsss: Em-
powering the downlink communication for backscatter systems,” in Proceedings of USENIX
NSDI, 2022.

[208] X. Guo, L. Shangguan, Y. He, N. Jing, J. Zhang, H. Jiang, and Y. Liu, “Saiyan: Design and
implementation of a low-power demodulator for lora backscatter systems,” in Proceedings
of USENIX NSDI, 2022.

[209] P. Radoglou-Grammatikis, P. Sarigiannidis, T. Lagkas, and I. Moscholios, “A compilation of
uav applications for precision agriculture,” Computer Networks, vol. 172, p. 107148, 2020.

170

[210] Unlicensed White Space Device Operations in the Television Bands, Federal Communica-

tions Commission, 2023.

[211] J. Du, Y. Liu, Y. Ren, L. Liu, and Z. Cao, “Loratrimmer: Optimal energy condensation with

chirp trimming for lora weak signal decoding,” in Proceedings of ACM MobiCom, 2024.

[212] E. W. Weisstein, “Heaviside step function,” https://mathworld. wolfram. com/, 2002.

[213] G. S. GADGETS, “Hackrf one,” https://greatscottgadgets.com/hackrf/one/, accessed 06-

Jun-2025.

[214] Y. Sangar, Y. Biradavolu, K. Pederson, V. Ranganathan, and B. Krishnaswamy, “Pact:
Scalable, long-range communication for monitoring and tracking systems using battery-less
tags,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technolo-
gies, vol. 6, no. 4, pp. 1–27, 2023.

[215] S. Tong, Z. Xu, and J. Wang, “Colora: Enabling multi-packet reception in lora,” in Proceed-

ings of IEEE INFOCOM, 2020.

[216] X. Xia, N. Hou, Y. Zheng, and T. Gu, “Pcube: scaling lora concurrent transmissions with
reception diversities,” ACM Transactions on Sensor Networks, vol. 18, no. 4, pp. 1–25, 2023.

[217] J. Luo, Z. Xu, J. Lin, C. Chen, and R. Xiong, “Ch-mac: Achieving low-latency reliable
communication via coding and hopping in lpwan,” ACM Transactions on Internet of Things,
vol. 4, no. 4, pp. 1–25, 2023.

[218] G. Yang, R. Dai, and Y.-C. Liang, “Energy-efficient uav backscatter communication with
joint trajectory design and resource optimization,” IEEE Transactions on Wireless Commu-
nications, vol. 20, no. 2, pp. 926–941, 2020.

[219] S. Yang, Y. Deng, X. Tang, Y. Ding, and J. Zhou, “Energy efficiency optimization for uav-
assisted backscatter communications,” IEEE Communications Letters, vol. 23, no. 11, pp.
2041–2045, 2019.

[220] R. Han, L. Bai, Y. Wen, J. Liu, J. Choi, and W. Zhang, “Uav-aided backscatter communica-
tions: Performance analysis and trajectory optimization,” IEEE Journal on Selected Areas
in Communications, vol. 39, no. 10, pp. 3129–3143, 2021.

[221] “Lora by the numbers,” https://www.semtech.com/lora, accessed 06-Jun-2025.

[222] “Amazon

sidewalk,”

https://www.amazon.com/Amazon-Sidewalk/b?ie=UTF8&node=

21328123011, accessed 06-Jun-2025.

[223] “Microsoft

farmbeats,”
farmbeats-iot-agriculture/, accessed 06-Jun-2025.

https://www.microsoft.com/en-us/research/project/

[224] T. Chakraborty, H. Shi, Z. Kapetanovic, B. Priyantha, D. Vasisht, B. Vu, P. Pandit, P. Pillai,
Y. Chabria, A. Nelson, M. Daum, and R. Chandra, “Whisper: Iot in the tv white space
spectrum,” in Proceedings of USENIX NSDI, 2022.

171

[225] Y. Yao, Z. Ma, and Z. Cao, “Losee: Long-range shared bike communication system based

on lorawan protocol.” in Proceedings of EWSN, 2019, pp. 407–412.

[226] L. Liu, Y. Yao, Z. Cao, and M. Zhang, “Deeplora: Learning accurate path loss model for

long distance links in lpwan,” in Proceedings of IEEE INFOCOM, 2021.

[227] Semtech, “Sx1276/77/78/79 datasheet,” https://www.semtech.com/products/wireless-rf/

lora-transceivers, accessed 06-Jun-2025.

[228] B. Ghena, J. Adkins, L. Shangguan, K. Jamieson, P. Levis, and P. Dutta, “Challenge: Un-
licensed lpwans are not yet the path to ubiquitous connectivity,” in Proceedings of ACM
MobiCom, 2019.

[229] M. O. Shahid, M. Philipose, K. Chintalapudi, S. Banerjee, and B. Krishnaswamy, “Concur-
rent interference cancellation: Decoding multi-packet collisions in lora,” in Proceedings of
ACM SIGCOMM, 2021.

[230] D. Mu, Y. Chen, J. Shi, and M. Sha, “Runtime control of lora spreading factor for campus

shuttle monitoring,” in Proceedings of IEEE ICNP, 2020.

[231] J. Yang, Z. Xu, and J. Wang, “Ferrylink: Combating link degradation for practical lpwan

deployments,” in Proceedings of IEEE ICPADS, 2021.

[232] A. Gadre, R. Narayanan, A. Luong, A. Rowe, B. Iannucci, and S. Kumar, “Frequency
configuration for low-power wide-area networks in heartbeat,” in Proceedings of USENIX
NSDI, 2020.

[233] N. Hou, X. Xia, and Y. Zheng, “Don’t miss weak packets: Boosting lora reception with

antenna diversities,” in Proceedings of IEEE INFOCOM, 2022.

[234] A. Balanuta, N. Pereira, S. Kumar, and A. Rowe, “A cloud-optimized link layer for low-power

wide-area networks,” in Proceedings of ACM MobiSys, 2020.

[235] X. Xia, Q. Chen, N. Hou, Y. Zheng, and M. Li, “Xcopy: Boosting weak links for reliable

lora communication,” in Proceedings of ACM MobiCom, 2023.

[236] S. Tong, Z. Shen, Y. Liu, and J. Wang, “Combating link dynamics for reliable lora connection

in urban settings,” in Proceedings of ACM MobiCom, 2021.

[237] S. Yu, X. Xia, N. Hou, Y. Zheng, and T. Gu, “Revolutionizing lora gateway with xgate:
Scalable concurrent transmission across massive logical channels,” in Proceedings of ACM
MobiCom, 2024.

[238] S. Tong, J. Wang, and Y. Liu, “Combating packet collisions using non-stationary signal scal-

ing in LPWANs,” in Proceedings of ACM MobiSys, 2020.

[239] S. Tong, Z. Xu, and J. Wang, “CoLoRa: Enabling Multi-Packet Reception in LoRa,” in

Proceedings of IEEE INFOCOM, 2020.

172

[240] X. Xia, Y. Zheng, and T. Gu, “FTrack: parallel decoding for LoRa transmissions,” in Pro-

ceedings of ACM SenSys, 2019.

[241] R. Eletreby, D. Zhang, S. Kumar, and O. Yağan, “Empowering Low-Power Wide Area Net-

works in Urban Settings,” in Proceedings of ACM SIGCOMM, 2017.

[242] F. Yu, X. Zheng, L. Liu, and H. Ma, “Enabling concurrency for non-orthogonal lora chan-

nels,” in Proceedings of ACM MobiCom, 2023.

[243] Y. Li, J. Yang, and J. Wang, “DyLoRa: Towards Energy Efficient Dynamic LoRa Transmis-

sion Control,” in Proceedings of IEEE INFOCOM, 2020.

[244] W. Gao, Z. Zhao, and G. Min, “AdapLoRa: Resource Adaptation for Maximizing Network

Lifetime in LoRa networks,” in Proceedings of IEEE ICNP, 2020.

[245] Z. Sun, T. Ni, H. Yang, K. Liu, Y. Zhang, T. Gu, and W. Xu, “Flora: Energy-efficient, reliable,
and beamforming-assisted over-the-air firmware update in lora networks,” in Proceedings of
ACM/IEEE IPSN, 2023.

[246] K. Yang and W. Du, “Lldpc: A low-density parity-check coding scheme for lora networks,”

in Proceedings of ACM Sensys, 2022.

[247] C. Li, Z. Cao, and L. Xiao, “Curvealoha: Non-linear chirps enabled high throughput random

channel access for lora,” in Proceedings of IEEE INFOCOM, 2022.

[248] T. Elshabrawy and J. Robert, “Interleaved chirp spreading lora-based modulation,” IEEE

Internet of Things Journal, vol. 6, no. 2, pp. 3855–3863, 2019.

[249] “Monsoon solutions

inc. high voltage power monitor,”

https://www.msoon.com/

high-voltage-power-monitor, accessed 06-Jun-2025.

[250] R. PI, “Raspberry pi 4,” https://www.raspberrypi.com/products/raspberry-pi-4-model-b/,

Retrieved by Mar 28 2024.

[251] Z. Xu, P. Xie, J. Wang, and Y. Liu, “Ostinato: Combating lora weak links in real deploy-

ments,” in Proceedings of IEEE ICNP, 2022.

[252] A. Gamage, C. J. Liando, C. Gu, R. Tan, and M. Li, “Lmac:efficient carrier-sense multiple

access for lora,” in Proceedings of ACM MobiCom, 2020.

[253] Semtech, “Lorawan 1.0.4 specification in depth for end device developers,” https://learn.

semtech.com/course/view.php?id=6&section=1, accessed 06-Jun-2025.

[254] M. Slabicki, G. Premsankar, and M. Di Francesco, “Adaptive configuration of lora networks
for dense iot deployments,” in NOMS 2018-2018 IEEE/IFIP Network Operations and Man-
agement Symposium.

IEEE, 2018, pp. 1–9.

[255] Semtech, “Implementing adaptive data rate,” https://learn.semtech.com/mod/book/view.

php?id=174, accessed 06-Jun-2025.

173